*
* src/backend/optimizer/path/allpaths.c
* set_append_rel_pathlist()
+ * generate_mergeappend_paths()
+ * get_cheapest_parameterized_child_path()
* accumulate_append_subpath()
- * set_dummy_rel_pathlist()
* standard_join_search()
*
* src/backend/optimizer/path/joinrels.c
* mark_dummy_rel()
* restriction_is_constant_false()
*
- * Portions Copyright (c) 1996-2011, PostgreSQL Global Development Group
+ *
+ * Portions Copyright (c) 1996-2019, PostgreSQL Global Development Group
* Portions Copyright (c) 1994, Regents of the University of California
*
*-------------------------------------------------------------------------
/*
* set_append_rel_pathlist
* Build access paths for an "append relation"
- *
- * The passed-in rel and RTE represent the entire append relation. The
- * relation's contents are computed by appending together the output of
- * the individual member relations. Note that in the inheritance case,
- * the first member relation is actually the same table as is mentioned in
- * the parent RTE ... but it has a different RTE and RelOptInfo. This is
- * a good thing because their outputs are not the same size.
*/
static void
set_append_rel_pathlist(PlannerInfo *root, RelOptInfo *rel,
int parentRTindex = rti;
List *live_childrels = NIL;
List *subpaths = NIL;
+ bool subpaths_valid = true;
List *all_child_pathkeys = NIL;
- double parent_rows;
- double parent_size;
- double *parent_attrsizes;
- int nattrs;
+ List *all_child_outers = NIL;
ListCell *l;
/*
- * Initialize to compute size estimates for whole append relation.
- *
- * We handle width estimates by weighting the widths of different child
- * rels proportionally to their number of rows. This is sensible because
- * the use of width estimates is mainly to compute the total relation
- * "footprint" if we have to sort or hash it. To do this, we sum the
- * total equivalent size (in "double" arithmetic) and then divide by the
- * total rowcount estimate. This is done separately for the total rel
- * width and each attribute.
- *
- * Note: if you consider changing this logic, beware that child rels could
- * have zero rows and/or width, if they were excluded by constraints.
- */
- parent_rows = 0;
- parent_size = 0;
- nattrs = rel->max_attr - rel->min_attr + 1;
- parent_attrsizes = (double *) palloc0(nattrs * sizeof(double));
-
- /*
- * Generate access paths for each member relation, and pick the cheapest
- * path for each one.
+ * Generate access paths for each member relation, and remember the
+ * cheapest path for each one. Also, identify all pathkeys (orderings)
+ * and parameterizations (required_outer sets) available for the member
+ * relations.
*/
foreach(l, root->append_rel_list)
{
int childRTindex;
RangeTblEntry *childRTE;
RelOptInfo *childrel;
- List *childquals;
- Node *childqual;
ListCell *lcp;
- ListCell *parentvars;
- ListCell *childvars;
/* append_rel_list contains all append rels; ignore others */
if (appinfo->parent_relid != parentRTindex)
continue;
+ /* Re-locate the child RTE and RelOptInfo */
childRTindex = appinfo->child_relid;
childRTE = root->simple_rte_array[childRTindex];
+ childrel = root->simple_rel_array[childRTindex];
/*
- * The child rel's RelOptInfo was already created during
- * add_base_rels_to_query.
+ * Compute the child's access paths.
*/
- childrel = find_base_rel(root, childRTindex);
- Assert(childrel->reloptkind == RELOPT_OTHER_MEMBER_REL);
+ set_rel_pathlist(root, childrel, childRTindex, childRTE);
/*
- * We have to copy the parent's targetlist and quals to the child,
- * with appropriate substitution of variables. However, only the
- * baserestrictinfo quals are needed before we can check for
- * constraint exclusion; so do that first and then check to see if we
- * can disregard this child.
- *
- * As of 8.4, the child rel's targetlist might contain non-Var
- * expressions, which means that substitution into the quals could
- * produce opportunities for const-simplification, and perhaps even
- * pseudoconstant quals. To deal with this, we strip the RestrictInfo
- * nodes, do the substitution, do const-simplification, and then
- * reconstitute the RestrictInfo layer.
+ * If child is dummy, ignore it.
*/
- childquals = get_all_actual_clauses(rel->baserestrictinfo);
- childquals = (List *) adjust_appendrel_attrs((Node *) childquals,
- appinfo);
- childqual = eval_const_expressions(root, (Node *)
- make_ands_explicit(childquals));
- if (childqual && IsA(childqual, Const) &&
- (((Const *) childqual)->constisnull ||
- !DatumGetBool(((Const *) childqual)->constvalue)))
- {
- /*
- * Restriction reduces to constant FALSE or constant NULL after
- * substitution, so this child need not be scanned.
- */
- set_dummy_rel_pathlist(childrel);
- continue;
- }
- childquals = make_ands_implicit((Expr *) childqual);
- childquals = make_restrictinfos_from_actual_clauses(root,
- childquals);
- childrel->baserestrictinfo = childquals;
-
- if (relation_excluded_by_constraints(root, childrel, childRTE))
- {
- /*
- * This child need not be scanned, so we can omit it from the
- * appendrel. Mark it with a dummy cheapest-path though, in case
- * best_appendrel_indexscan() looks at it later.
- */
- set_dummy_rel_pathlist(childrel);
+ if (IS_DUMMY_REL(childrel))
continue;
- }
-
- /*
- * CE failed, so finish copying/modifying targetlist and join quals.
- *
- * Note: the resulting childrel->reltargetlist may contain arbitrary
- * expressions, which normally would not occur in a reltargetlist.
- * That is okay because nothing outside of this routine will look at
- * the child rel's reltargetlist. We do have to cope with the case
- * while constructing attr_widths estimates below, though.
- */
- childrel->joininfo = (List *)
- adjust_appendrel_attrs((Node *) rel->joininfo,
- appinfo);
- childrel->reltargetlist = (List *)
- adjust_appendrel_attrs((Node *) rel->reltargetlist,
- appinfo);
/*
- * We have to make child entries in the EquivalenceClass data
- * structures as well. This is needed either if the parent
- * participates in some eclass joins (because we will want to consider
- * inner-indexscan joins on the individual children) or if the parent
- * has useful pathkeys (because we should try to build MergeAppend
- * paths that produce those sort orderings).
+ * Child is live, so add it to the live_childrels list for use below.
*/
- if (rel->has_eclass_joins || has_useful_pathkeys(root, rel))
- add_child_rel_equivalences(root, appinfo, rel, childrel);
- childrel->has_eclass_joins = rel->has_eclass_joins;
-
- /*
- * Note: we could compute appropriate attr_needed data for the child's
- * variables, by transforming the parent's attr_needed through the
- * translated_vars mapping. However, currently there's no need
- * because attr_needed is only examined for base relations not
- * otherrels. So we just leave the child's attr_needed empty.
- */
-
- /* Remember which childrels are live, for MergeAppend logic below */
live_childrels = lappend(live_childrels, childrel);
/*
- * Compute the child's access paths, and add the cheapest one to the
- * Append path we are constructing for the parent.
+ * If child has an unparameterized cheapest-total path, add that to
+ * the unparameterized Append path we are constructing for the parent.
+ * If not, there's no workable unparameterized path.
*/
- set_rel_pathlist(root, childrel, childRTindex, childRTE);
-
- subpaths = accumulate_append_subpath(subpaths,
- childrel->cheapest_total_path);
+ if (childrel->cheapest_total_path->param_info == NULL)
+ subpaths = accumulate_append_subpath(subpaths,
+ childrel->cheapest_total_path);
+ else
+ subpaths_valid = false;
/*
- * Collect a list of all the available path orderings for all the
- * children. We use this as a heuristic to indicate which sort
- * orderings we should build MergeAppend paths for.
+ * Collect lists of all the available path orderings and
+ * parameterizations for all the children. We use these as a
+ * heuristic to indicate which sort orderings and parameterizations we
+ * should build Append and MergeAppend paths for.
*/
foreach(lcp, childrel->pathlist)
{
Path *childpath = (Path *) lfirst(lcp);
List *childkeys = childpath->pathkeys;
- ListCell *lpk;
- bool found = false;
-
- /* Ignore unsorted paths */
- if (childkeys == NIL)
- continue;
+ Relids childouter = PATH_REQ_OUTER(childpath);
- /* Have we already seen this ordering? */
- foreach(lpk, all_child_pathkeys)
+ /* Unsorted paths don't contribute to pathkey list */
+ if (childkeys != NIL)
{
- List *existing_pathkeys = (List *) lfirst(lpk);
+ ListCell *lpk;
+ bool found = false;
- if (compare_pathkeys(existing_pathkeys,
- childkeys) == PATHKEYS_EQUAL)
+ /* Have we already seen this ordering? */
+ foreach(lpk, all_child_pathkeys)
{
- found = true;
- break;
+ List *existing_pathkeys = (List *) lfirst(lpk);
+
+ if (compare_pathkeys(existing_pathkeys,
+ childkeys) == PATHKEYS_EQUAL)
+ {
+ found = true;
+ break;
+ }
+ }
+ if (!found)
+ {
+ /* No, so add it to all_child_pathkeys */
+ all_child_pathkeys = lappend(all_child_pathkeys,
+ childkeys);
}
}
- if (!found)
- {
- /* No, so add it to all_child_pathkeys */
- all_child_pathkeys = lappend(all_child_pathkeys, childkeys);
- }
- }
- /*
- * Accumulate size information from each child.
- */
- if (childrel->rows > 0)
- {
- parent_rows += childrel->rows;
- parent_size += childrel->width * childrel->rows;
-
- /*
- * Accumulate per-column estimates too. We need not do anything
- * for PlaceHolderVars in the parent list. If child expression
- * isn't a Var, or we didn't record a width estimate for it, we
- * have to fall back on a datatype-based estimate.
- *
- * By construction, child's reltargetlist is 1-to-1 with parent's.
- */
- forboth(parentvars, rel->reltargetlist,
- childvars, childrel->reltargetlist)
+ /* Unparameterized paths don't contribute to param-set list */
+ if (childouter)
{
- Var *parentvar = (Var *) lfirst(parentvars);
- Node *childvar = (Node *) lfirst(childvars);
+ ListCell *lco;
+ bool found = false;
- if (IsA(parentvar, Var))
+ /* Have we already seen this param set? */
+ foreach(lco, all_child_outers)
{
- int pndx = parentvar->varattno - rel->min_attr;
- int32 child_width = 0;
+ Relids existing_outers = (Relids) lfirst(lco);
- if (IsA(childvar, Var))
+ if (bms_equal(existing_outers, childouter))
{
- int cndx = ((Var *) childvar)->varattno - childrel->min_attr;
-
- child_width = childrel->attr_widths[cndx];
+ found = true;
+ break;
}
- if (child_width <= 0)
- child_width = get_typavgwidth(exprType(childvar),
- exprTypmod(childvar));
- Assert(child_width > 0);
- parent_attrsizes[pndx] += child_width * childrel->rows;
+ }
+ if (!found)
+ {
+ /* No, so add it to all_child_outers */
+ all_child_outers = lappend(all_child_outers,
+ childouter);
}
}
}
}
/*
- * Save the finished size estimates.
+ * If we found unparameterized paths for all children, build an unordered,
+ * unparameterized Append path for the rel. (Note: this is correct even
+ * if we have zero or one live subpath due to constraint exclusion.)
*/
- rel->rows = parent_rows;
- if (parent_rows > 0)
- {
- int i;
-
- rel->width = rint(parent_size / parent_rows);
- for (i = 0; i < nattrs; i++)
- rel->attr_widths[i] = rint(parent_attrsizes[i] / parent_rows);
- }
- else
- rel->width = 0; /* attr_widths should be zero already */
+ if (subpaths_valid)
+ add_path(rel, (Path *) create_append_path(rel, subpaths, NULL));
/*
- * Set "raw tuples" count equal to "rows" for the appendrel; needed
- * because some places assume rel->tuples is valid for any baserel.
+ * Also build unparameterized MergeAppend paths based on the collected
+ * list of child pathkeys.
*/
- rel->tuples = parent_rows;
-
- pfree(parent_attrsizes);
+ if (subpaths_valid)
+ generate_mergeappend_paths(root, rel, live_childrels,
+ all_child_pathkeys);
/*
- * Next, build an unordered Append path for the rel. (Note: this is
- * correct even if we have zero or one live subpath due to constraint
- * exclusion.)
+ * Build Append paths for each parameterization seen among the child rels.
+ * (This may look pretty expensive, but in most cases of practical
+ * interest, the child rels will expose mostly the same parameterizations,
+ * so that not that many cases actually get considered here.)
+ *
+ * The Append node itself cannot enforce quals, so all qual checking must
+ * be done in the child paths. This means that to have a parameterized
+ * Append path, we must have the exact same parameterization for each
+ * child path; otherwise some children might be failing to check the
+ * moved-down quals. To make them match up, we can try to increase the
+ * parameterization of lesser-parameterized paths.
*/
- add_path(rel, (Path *) create_append_path(rel, subpaths));
+ foreach(l, all_child_outers)
+ {
+ Relids required_outer = (Relids) lfirst(l);
+ ListCell *lcr;
- /*
- * Next, build MergeAppend paths based on the collected list of child
- * pathkeys. We consider both cheapest-startup and cheapest-total cases,
- * ie, for each interesting ordering, collect all the cheapest startup
- * subpaths and all the cheapest total paths, and build a MergeAppend path
- * for each list.
- */
- foreach(l, all_child_pathkeys)
+ /* Select the child paths for an Append with this parameterization */
+ subpaths = NIL;
+ subpaths_valid = true;
+ foreach(lcr, live_childrels)
+ {
+ RelOptInfo *childrel = (RelOptInfo *) lfirst(lcr);
+ Path *subpath;
+
+ subpath = get_cheapest_parameterized_child_path(root,
+ childrel,
+ required_outer);
+ if (subpath == NULL)
+ {
+ /* failed to make a suitable path for this child */
+ subpaths_valid = false;
+ break;
+ }
+ subpaths = accumulate_append_subpath(subpaths, subpath);
+ }
+
+ if (subpaths_valid)
+ add_path(rel, (Path *)
+ create_append_path(rel, subpaths, required_outer));
+ }
+}
+
+/*
+ * generate_mergeappend_paths
+ * Generate MergeAppend paths for an append relation
+ *
+ * Generate a path for each ordering (pathkey list) appearing in
+ * all_child_pathkeys.
+ *
+ * We consider both cheapest-startup and cheapest-total cases, ie, for each
+ * interesting ordering, collect all the cheapest startup subpaths and all the
+ * cheapest total paths, and build a MergeAppend path for each case.
+ *
+ * We don't currently generate any parameterized MergeAppend paths. While
+ * it would not take much more code here to do so, it's very unclear that it
+ * is worth the planning cycles to investigate such paths: there's little
+ * use for an ordered path on the inside of a nestloop. In fact, it's likely
+ * that the current coding of add_path would reject such paths out of hand,
+ * because add_path gives no credit for sort ordering of parameterized paths,
+ * and a parameterized MergeAppend is going to be more expensive than the
+ * corresponding parameterized Append path. If we ever try harder to support
+ * parameterized mergejoin plans, it might be worth adding support for
+ * parameterized MergeAppends to feed such joins. (See notes in
+ * optimizer/README for why that might not ever happen, though.)
+ */
+static void
+generate_mergeappend_paths(PlannerInfo *root, RelOptInfo *rel,
+ List *live_childrels,
+ List *all_child_pathkeys)
+{
+ ListCell *lcp;
+
+ foreach(lcp, all_child_pathkeys)
{
- List *pathkeys = (List *) lfirst(l);
+ List *pathkeys = (List *) lfirst(lcp);
List *startup_subpaths = NIL;
List *total_subpaths = NIL;
bool startup_neq_total = false;
cheapest_startup =
get_cheapest_path_for_pathkeys(childrel->pathlist,
pathkeys,
+ NULL,
STARTUP_COST);
cheapest_total =
get_cheapest_path_for_pathkeys(childrel->pathlist,
pathkeys,
+ NULL,
TOTAL_COST);
/*
- * If we can't find any paths with the right order just add the
- * cheapest-total path; we'll have to sort it.
+ * If we can't find any paths with the right order just use the
+ * cheapest-total path; we'll have to sort it later.
*/
- if (cheapest_startup == NULL)
- cheapest_startup = childrel->cheapest_total_path;
- if (cheapest_total == NULL)
- cheapest_total = childrel->cheapest_total_path;
+ if (cheapest_startup == NULL || cheapest_total == NULL)
+ {
+ cheapest_startup = cheapest_total =
+ childrel->cheapest_total_path;
+ /* Assert we do have an unparameterized path for this child */
+ Assert(cheapest_total->param_info == NULL);
+ }
/*
* Notice whether we actually have different paths for the
add_path(rel, (Path *) create_merge_append_path(root,
rel,
startup_subpaths,
- pathkeys));
+ pathkeys,
+ NULL));
if (startup_neq_total)
add_path(rel, (Path *) create_merge_append_path(root,
rel,
total_subpaths,
- pathkeys));
+ pathkeys,
+ NULL));
}
+}
- /* Select cheapest path */
- set_cheapest(rel);
+/*
+ * get_cheapest_parameterized_child_path
+ * Get cheapest path for this relation that has exactly the requested
+ * parameterization.
+ *
+ * Returns NULL if unable to create such a path.
+ */
+static Path *
+get_cheapest_parameterized_child_path(PlannerInfo *root, RelOptInfo *rel,
+ Relids required_outer)
+{
+ Path *cheapest;
+ ListCell *lc;
+
+ /*
+ * Look up the cheapest existing path with no more than the needed
+ * parameterization. If it has exactly the needed parameterization, we're
+ * done.
+ */
+ cheapest = get_cheapest_path_for_pathkeys(rel->pathlist,
+ NIL,
+ required_outer,
+ TOTAL_COST);
+ Assert(cheapest != NULL);
+ if (bms_equal(PATH_REQ_OUTER(cheapest), required_outer))
+ return cheapest;
+
+ /*
+ * Otherwise, we can "reparameterize" an existing path to match the given
+ * parameterization, which effectively means pushing down additional
+ * joinquals to be checked within the path's scan. However, some existing
+ * paths might check the available joinquals already while others don't;
+ * therefore, it's not clear which existing path will be cheapest after
+ * reparameterization. We have to go through them all and find out.
+ */
+ cheapest = NULL;
+ foreach(lc, rel->pathlist)
+ {
+ Path *path = (Path *) lfirst(lc);
+
+ /* Can't use it if it needs more than requested parameterization */
+ if (!bms_is_subset(PATH_REQ_OUTER(path), required_outer))
+ continue;
+
+ /*
+ * Reparameterization can only increase the path's cost, so if it's
+ * already more expensive than the current cheapest, forget it.
+ */
+ if (cheapest != NULL &&
+ compare_path_costs(cheapest, path, TOTAL_COST) <= 0)
+ continue;
+
+ /* Reparameterize if needed, then recheck cost */
+ if (!bms_equal(PATH_REQ_OUTER(path), required_outer))
+ {
+ path = reparameterize_path(root, path, required_outer, 1.0);
+ if (path == NULL)
+ continue; /* failed to reparameterize this one */
+ Assert(bms_equal(PATH_REQ_OUTER(path), required_outer));
+
+ if (cheapest != NULL &&
+ compare_path_costs(cheapest, path, TOTAL_COST) <= 0)
+ continue;
+ }
+
+ /* We have a new best path */
+ cheapest = path;
+ }
+
+ /* Return the best path, or NULL if we found no suitable candidate */
+ return cheapest;
}
/*
* accumulate_append_subpath
* Add a subpath to the list being built for an Append or MergeAppend
*
- * It's possible that the child is itself an Append path, in which case
- * we can "cut out the middleman" and just add its child paths to our
- * own list. (We don't try to do this earlier because we need to
- * apply both levels of transformation to the quals.)
+ * It's possible that the child is itself an Append or MergeAppend path, in
+ * which case we can "cut out the middleman" and just add its child paths to
+ * our own list. (We don't try to do this earlier because we need to apply
+ * both levels of transformation to the quals.)
+ *
+ * Note that if we omit a child MergeAppend in this way, we are effectively
+ * omitting a sort step, which seems fine: if the parent is to be an Append,
+ * its result would be unsorted anyway, while if the parent is to be a
+ * MergeAppend, there's no point in a separate sort on a child.
*/
static List *
accumulate_append_subpath(List *subpaths, Path *path)
/* list_copy is important here to avoid sharing list substructure */
return list_concat(subpaths, list_copy(apath->subpaths));
}
+ else if (IsA(path, MergeAppendPath))
+ {
+ MergeAppendPath *mpath = (MergeAppendPath *) path;
+
+ /* list_copy is important here to avoid sharing list substructure */
+ return list_concat(subpaths, list_copy(mpath->subpaths));
+ }
else
return lappend(subpaths, path);
}
/*
- * set_dummy_rel_pathlist
- * Build a dummy path for a relation that's been excluded by constraints
- *
- * Rather than inventing a special "dummy" path type, we represent this as an
- * AppendPath with no members (see also IS_DUMMY_PATH macro).
- */
-static void
-set_dummy_rel_pathlist(RelOptInfo *rel)
-{
- /* Set dummy size estimates --- we leave attr_widths[] as zeroes */
- rel->rows = 0;
- rel->width = 0;
-
- add_path(rel, (Path *) create_append_path(rel, NIL));
-
- /* Select cheapest path (pretty easy in this case...) */
- set_cheapest(rel);
-}
-
-/*
* standard_join_search
* Find possible joinpaths for a query by successively finding ways
* to join component relations into join relations.
* independent jointree items in the query. This is > 1.
*
* 'initial_rels' is a list of RelOptInfo nodes for each independent
- * jointree item. These are the components to be joined together.
+ * jointree item. These are the components to be joined together.
* Note that levels_needed == list_length(initial_rels).
*
* Returns the final level of join relations, i.e., the relation that is
* needed for these paths need have been instantiated.
*
* Note to plugin authors: the functions invoked during standard_join_search()
- * modify root->join_rel_list and root->join_rel_hash. If you want to do more
+ * modify root->join_rel_list and root->join_rel_hash. If you want to do more
* than one join-order search, you'll probably need to save and restore the
* original states of those data structures. See geqo_eval() for an example.
*/
* We prefer to join using join clauses, but if we find a rel of level-1
* members that has no join clauses, we will generate Cartesian-product
* joins against all initial rels not already contained in it.
- *
- * In the first pass (level == 2), we try to join each initial rel to each
- * initial rel that appears later in joinrels[1]. (The mirror-image joins
- * are handled automatically by make_join_rel.) In later passes, we try
- * to join rels of size level-1 from joinrels[level-1] to each initial rel
- * in joinrels[1].
*/
foreach(r, joinrels[level - 1])
{
RelOptInfo *old_rel = (RelOptInfo *) lfirst(r);
- ListCell *other_rels;
-
- if (level == 2)
- other_rels = lnext(r); /* only consider remaining initial
- * rels */
- else
- other_rels = list_head(joinrels[1]); /* consider all initial
- * rels */
if (old_rel->joininfo != NIL || old_rel->has_eclass_joins ||
has_join_restriction(root, old_rel))
{
/*
- * Note that if all available join clauses for this rel require
- * more than one other rel, we will fail to make any joins against
- * it here. In most cases that's OK; it'll be considered by
- * "bushy plan" join code in a higher-level pass where we have
- * those other rels collected into a join rel.
+ * There are join clauses or join order restrictions relevant to
+ * this rel, so consider joins between this rel and (only) those
+ * initial rels it is linked to by a clause or restriction.
*
- * See also the last-ditch case below.
+ * At level 2 this condition is symmetric, so there is no need to
+ * look at initial rels before this one in the list; we already
+ * considered such joins when we were at the earlier rel. (The
+ * mirror-image joins are handled automatically by make_join_rel.)
+ * In later passes (level > 2), we join rels of the previous level
+ * to each initial rel they don't already include but have a join
+ * clause or restriction with.
*/
+ ListCell *other_rels;
+
+ if (level == 2) /* consider remaining initial rels */
+ other_rels = lnext(r);
+ else /* consider all initial rels */
+ other_rels = list_head(joinrels[1]);
+
make_rels_by_clause_joins(root,
old_rel,
other_rels);
* Oops, we have a relation that is not joined to any other
* relation, either directly or by join-order restrictions.
* Cartesian product time.
+ *
+ * We consider a cartesian product with each not-already-included
+ * initial rel, whether it has other join clauses or not. At
+ * level 2, if there are two or more clauseless initial rels, we
+ * will redundantly consider joining them in both directions; but
+ * such cases aren't common enough to justify adding complexity to
+ * avoid the duplicated effort.
*/
make_rels_by_clauseless_joins(root,
old_rel,
- other_rels);
+ list_head(joinrels[1]));
}
}
ListCell *r2;
/*
- * We can ignore clauseless joins here, *except* when they
+ * We can ignore relations without join clauses here, unless they
* participate in join-order restrictions --- then we might have
* to force a bushy join plan.
*/
{
/*
* OK, we can build a rel of the right level from this
- * pair of rels. Do so if there is at least one usable
- * join clause or a relevant join restriction.
+ * pair of rels. Do so if there is at least one relevant
+ * join clause or join order restriction.
*/
if (have_relevant_joinclause(root, old_rel, new_rel) ||
have_join_order_restriction(root, old_rel, new_rel))
}
}
- /*
+ /*----------
* Last-ditch effort: if we failed to find any usable joins so far, force
* a set of cartesian-product joins to be generated. This handles the
* special case where all the available rels have join clauses but we
- * cannot use any of those clauses yet. An example is
+ * cannot use any of those clauses yet. This can only happen when we are
+ * considering a join sub-problem (a sub-joinlist) and all the rels in the
+ * sub-problem have only join clauses with rels outside the sub-problem.
+ * An example is
*
- * SELECT * FROM a,b,c WHERE (a.f1 + b.f2 + c.f3) = 0;
+ * SELECT ... FROM a INNER JOIN b ON TRUE, c, d, ...
+ * WHERE a.w = c.x and b.y = d.z;
*
- * The join clause will be usable at level 3, but at level 2 we have no
- * choice but to make cartesian joins. We consider only left-sided and
- * right-sided cartesian joins in this case (no bushy).
+ * If the "a INNER JOIN b" sub-problem does not get flattened into the
+ * upper level, we must be willing to make a cartesian join of a and b;
+ * but the code above will not have done so, because it thought that both
+ * a and b have joinclauses. We consider only left-sided and right-sided
+ * cartesian joins in this case (no bushy).
+ *----------
*/
if (joinrels[level] == NIL)
{
foreach(r, joinrels[level - 1])
{
RelOptInfo *old_rel = (RelOptInfo *) lfirst(r);
- ListCell *other_rels;
-
- if (level == 2)
- other_rels = lnext(r); /* only consider remaining initial
- * rels */
- else
- other_rels = list_head(joinrels[1]); /* consider all initial
- * rels */
make_rels_by_clauseless_joins(root,
old_rel,
- other_rels);
+ list_head(joinrels[1]));
}
/*----------
* When special joins are involved, there may be no legal way
- * to make an N-way join for some values of N. For example consider
+ * to make an N-way join for some values of N. For example consider
*
* SELECT ... FROM t1 WHERE
* x IN (SELECT ... FROM t2,t3 WHERE ...) AND
* to accept failure at level 4 and go on to discover a workable
* bushy plan at level 5.
*
- * However, if there are no special joins then join_is_legal() should
- * never fail, and so the following sanity check is useful.
+ * However, if there are no special joins and no lateral references
+ * then join_is_legal() should never fail, and so the following sanity
+ * check is useful.
*----------
*/
- if (joinrels[level] == NIL && root->join_info_list == NIL)
+ if (joinrels[level] == NIL &&
+ root->join_info_list == NIL &&
+ !root->hasLateralRTEs)
elog(ERROR, "failed to build any %d-way joins", level);
}
}
SpecialJoinInfo *match_sjinfo;
bool reversed;
bool unique_ified;
- bool is_valid_inner;
+ bool must_be_leftjoin;
ListCell *l;
/*
- * Ensure output params are set on failure return. This is just to
+ * Ensure output params are set on failure return. This is just to
* suppress uninitialized-variable warnings from overly anal compilers.
*/
*sjinfo_p = NULL;
/*
* If we have any special joins, the proposed join might be illegal; and
- * in any case we have to determine its join type. Scan the join info
- * list for conflicts.
+ * in any case we have to determine its join type. Scan the join info
+ * list for matches and conflicts.
*/
match_sjinfo = NULL;
reversed = false;
unique_ified = false;
- is_valid_inner = true;
+ must_be_leftjoin = false;
foreach(l, root->join_info_list)
{
* If one input contains min_lefthand and the other contains
* min_righthand, then we can perform the SJ at this join.
*
- * Barf if we get matches to more than one SJ (is that possible?)
+ * Reject if we get matches to more than one SJ; that implies we're
+ * considering something that's not really valid.
*/
if (bms_is_subset(sjinfo->min_lefthand, rel1->relids) &&
bms_is_subset(sjinfo->min_righthand, rel2->relids))
}
else
{
- /*----------
- * Otherwise, the proposed join overlaps the RHS but isn't
- * a valid implementation of this SJ. It might still be
- * a legal join, however. If both inputs overlap the RHS,
- * assume that it's OK. Since the inputs presumably got past
- * this function's checks previously, they can't overlap the
- * LHS and their violations of the RHS boundary must represent
- * SJs that have been determined to commute with this one.
- * We have to allow this to work correctly in cases like
- * (a LEFT JOIN (b JOIN (c LEFT JOIN d)))
- * when the c/d join has been determined to commute with the join
- * to a, and hence d is not part of min_righthand for the upper
- * join. It should be legal to join b to c/d but this will appear
- * as a violation of the upper join's RHS.
- * Furthermore, if one input overlaps the RHS and the other does
- * not, we should still allow the join if it is a valid
- * implementation of some other SJ. We have to allow this to
- * support the associative identity
- * (a LJ b on Pab) LJ c ON Pbc = a LJ (b LJ c ON Pbc) on Pab
- * since joining B directly to C violates the lower SJ's RHS.
- * We assume that make_outerjoininfo() set things up correctly
- * so that we'll only match to some SJ if the join is valid.
- * Set flag here to check at bottom of loop.
- *----------
+ /*
+ * Otherwise, the proposed join overlaps the RHS but isn't a valid
+ * implementation of this SJ. But don't panic quite yet: the RHS
+ * violation might have occurred previously, in one or both input
+ * relations, in which case we must have previously decided that
+ * it was OK to commute some other SJ with this one. If we need
+ * to perform this join to finish building up the RHS, rejecting
+ * it could lead to not finding any plan at all. (This can occur
+ * because of the heuristics elsewhere in this file that postpone
+ * clauseless joins: we might not consider doing a clauseless join
+ * within the RHS until after we've performed other, validly
+ * commutable SJs with one or both sides of the clauseless join.)
+ * This consideration boils down to the rule that if both inputs
+ * overlap the RHS, we can allow the join --- they are either
+ * fully within the RHS, or represent previously-allowed joins to
+ * rels outside it.
*/
- if (sjinfo->jointype != JOIN_SEMI &&
- bms_overlap(rel1->relids, sjinfo->min_righthand) &&
+ if (bms_overlap(rel1->relids, sjinfo->min_righthand) &&
bms_overlap(rel2->relids, sjinfo->min_righthand))
- {
- /* seems OK */
- Assert(!bms_overlap(joinrelids, sjinfo->min_lefthand));
- }
- else
- is_valid_inner = false;
+ continue; /* assume valid previous violation of RHS */
+
+ /*
+ * The proposed join could still be legal, but only if we're
+ * allowed to associate it into the RHS of this SJ. That means
+ * this SJ must be a LEFT join (not SEMI or ANTI, and certainly
+ * not FULL) and the proposed join must not overlap the LHS.
+ */
+ if (sjinfo->jointype != JOIN_LEFT ||
+ bms_overlap(joinrelids, sjinfo->min_lefthand))
+ return false; /* invalid join path */
+
+ /*
+ * To be valid, the proposed join must be a LEFT join; otherwise
+ * it can't associate into this SJ's RHS. But we may not yet have
+ * found the SpecialJoinInfo matching the proposed join, so we
+ * can't test that yet. Remember the requirement for later.
+ */
+ must_be_leftjoin = true;
}
}
/*
- * Fail if violated some SJ's RHS and didn't match to another SJ. However,
- * "matching" to a semijoin we are implementing by unique-ification
- * doesn't count (think: it's really an inner join).
+ * Fail if violated any SJ's RHS and didn't match to a LEFT SJ: the
+ * proposed join can't associate into an SJ's RHS.
+ *
+ * Also, fail if the proposed join's predicate isn't strict; we're
+ * essentially checking to see if we can apply outer-join identity 3, and
+ * that's a requirement. (This check may be redundant with checks in
+ * make_outerjoininfo, but I'm not quite sure, and it's cheap to test.)
*/
- if (!is_valid_inner &&
- (match_sjinfo == NULL || unique_ified))
+ if (must_be_leftjoin &&
+ (match_sjinfo == NULL ||
+ match_sjinfo->jointype != JOIN_LEFT ||
+ !match_sjinfo->lhs_strict))
return false; /* invalid join path */
+ /*
+ * We also have to check for constraints imposed by LATERAL references.
+ */
+ if (root->hasLateralRTEs)
+ {
+ bool lateral_fwd;
+ bool lateral_rev;
+ Relids join_lateral_rels;
+
+ /*
+ * The proposed rels could each contain lateral references to the
+ * other, in which case the join is impossible. If there are lateral
+ * references in just one direction, then the join has to be done with
+ * a nestloop with the lateral referencer on the inside. If the join
+ * matches an SJ that cannot be implemented by such a nestloop, the
+ * join is impossible.
+ *
+ * Also, if the lateral reference is only indirect, we should reject
+ * the join; whatever rel(s) the reference chain goes through must be
+ * joined to first.
+ *
+ * Another case that might keep us from building a valid plan is the
+ * implementation restriction described by have_dangerous_phv().
+ */
+ lateral_fwd = bms_overlap(rel1->relids, rel2->lateral_relids);
+ lateral_rev = bms_overlap(rel2->relids, rel1->lateral_relids);
+ if (lateral_fwd && lateral_rev)
+ return false; /* have lateral refs in both directions */
+ if (lateral_fwd)
+ {
+ /* has to be implemented as nestloop with rel1 on left */
+ if (match_sjinfo &&
+ (reversed ||
+ unique_ified ||
+ match_sjinfo->jointype == JOIN_FULL))
+ return false; /* not implementable as nestloop */
+ /* check there is a direct reference from rel2 to rel1 */
+ if (!bms_overlap(rel1->relids, rel2->direct_lateral_relids))
+ return false; /* only indirect refs, so reject */
+ /* check we won't have a dangerous PHV */
+ if (have_dangerous_phv(root, rel1->relids, rel2->lateral_relids))
+ return false; /* might be unable to handle required PHV */
+ }
+ else if (lateral_rev)
+ {
+ /* has to be implemented as nestloop with rel2 on left */
+ if (match_sjinfo &&
+ (!reversed ||
+ unique_ified ||
+ match_sjinfo->jointype == JOIN_FULL))
+ return false; /* not implementable as nestloop */
+ /* check there is a direct reference from rel1 to rel2 */
+ if (!bms_overlap(rel2->relids, rel1->direct_lateral_relids))
+ return false; /* only indirect refs, so reject */
+ /* check we won't have a dangerous PHV */
+ if (have_dangerous_phv(root, rel2->relids, rel1->lateral_relids))
+ return false; /* might be unable to handle required PHV */
+ }
+
+ /*
+ * LATERAL references could also cause problems later on if we accept
+ * this join: if the join's minimum parameterization includes any rels
+ * that would have to be on the inside of an outer join with this join
+ * rel, then it's never going to be possible to build the complete
+ * query using this join. We should reject this join not only because
+ * it'll save work, but because if we don't, the clauseless-join
+ * heuristics might think that legality of this join means that some
+ * other join rel need not be formed, and that could lead to failure
+ * to find any plan at all. We have to consider not only rels that
+ * are directly on the inner side of an OJ with the joinrel, but also
+ * ones that are indirectly so, so search to find all such rels.
+ */
+ join_lateral_rels = min_join_parameterization(root, joinrelids,
+ rel1, rel2);
+ if (join_lateral_rels)
+ {
+ Relids join_plus_rhs = bms_copy(joinrelids);
+ bool more;
+
+ do
+ {
+ more = false;
+ foreach(l, root->join_info_list)
+ {
+ SpecialJoinInfo *sjinfo = (SpecialJoinInfo *) lfirst(l);
+
+ if (bms_overlap(sjinfo->min_lefthand, join_plus_rhs) &&
+ !bms_is_subset(sjinfo->min_righthand, join_plus_rhs))
+ {
+ join_plus_rhs = bms_add_members(join_plus_rhs,
+ sjinfo->min_righthand);
+ more = true;
+ }
+ /* full joins constrain both sides symmetrically */
+ if (sjinfo->jointype == JOIN_FULL &&
+ bms_overlap(sjinfo->min_righthand, join_plus_rhs) &&
+ !bms_is_subset(sjinfo->min_lefthand, join_plus_rhs))
+ {
+ join_plus_rhs = bms_add_members(join_plus_rhs,
+ sjinfo->min_lefthand);
+ more = true;
+ }
+ }
+ } while (more);
+ if (bms_overlap(join_plus_rhs, join_lateral_rels))
+ return false; /* will not be able to join to some RHS rel */
+ }
+ }
+
/* Otherwise, it's a valid join */
*sjinfo_p = match_sjinfo;
*reversed_p = reversed;
/*
* has_join_restriction
- * Detect whether the specified relation has join-order restrictions
- * due to being inside an outer join or an IN (sub-SELECT).
+ * Detect whether the specified relation has join-order restrictions,
+ * due to being inside an outer join or an IN (sub-SELECT),
+ * or participating in any LATERAL references or multi-rel PHVs.
*
* Essentially, this tests whether have_join_order_restriction() could
* succeed with this rel and some other one. It's OK if we sometimes
- * say "true" incorrectly. (Therefore, we don't bother with the relatively
+ * say "true" incorrectly. (Therefore, we don't bother with the relatively
* expensive has_legal_joinclause test.)
*/
static bool
{
ListCell *l;
+ if (rel->lateral_relids != NULL || rel->lateral_referencers != NULL)
+ return true;
+
+ foreach(l, root->placeholder_list)
+ {
+ PlaceHolderInfo *phinfo = (PlaceHolderInfo *) lfirst(l);
+
+ if (bms_is_subset(rel->relids, phinfo->ph_eval_at) &&
+ !bms_equal(rel->relids, phinfo->ph_eval_at))
+ return true;
+ }
+
foreach(l, root->join_info_list)
{
SpecialJoinInfo *sjinfo = (SpecialJoinInfo *) lfirst(l);
/*
* is_dummy_rel --- has relation been proven empty?
- *
- * If so, it will have a single path that is dummy.
*/
static bool
is_dummy_rel(RelOptInfo *rel)
{
- return (rel->cheapest_total_path != NULL &&
- IS_DUMMY_PATH(rel->cheapest_total_path));
+ return IS_DUMMY_REL(rel);
}
/*
* dummy.
*
* Also, when called during GEQO join planning, we are in a short-lived
- * memory context. We must make sure that the dummy path attached to a
+ * memory context. We must make sure that the dummy path attached to a
* baserel survives the GEQO cycle, else the baserel is trashed for future
* GEQO cycles. On the other hand, when we are marking a joinrel during GEQO,
* we don't want the dummy path to clutter the main planning context. Upshot
rel->pathlist = NIL;
/* Set up the dummy path */
- add_path(rel, (Path *) create_append_path(rel, NIL));
+ add_path(rel, (Path *) create_append_path(rel, NIL, NULL));
- /* Set or update cheapest_total_path */
+ /* Set or update cheapest_total_path and related fields */
set_cheapest(rel);
MemoryContextSwitchTo(oldcontext);
}
/*
- * restriction_is_constant_false --- is a restrictlist just FALSE?
+ * restriction_is_constant_false --- is a restrictlist just false?
*
- * In cases where a qual is provably constant FALSE, eval_const_expressions
+ * In cases where a qual is provably constant false, eval_const_expressions
* will generally have thrown away anything that's ANDed with it. In outer
* join situations this will leave us computing cartesian products only to
* decide there's no match for an outer row, which is pretty stupid. So,
* we need to detect the case.
*
- * If only_pushed_down is TRUE, then consider only pushed-down quals.
+ * If only_pushed_down is true, then consider only quals that are pushed-down
+ * from the point of view of the joinrel.
*/
static bool
-restriction_is_constant_false(List *restrictlist, bool only_pushed_down)
+restriction_is_constant_false(List *restrictlist,
+ RelOptInfo *joinrel,
+ bool only_pushed_down)
{
ListCell *lc;
RestrictInfo *rinfo = (RestrictInfo *) lfirst(lc);
Assert(IsA(rinfo, RestrictInfo));
- if (only_pushed_down && !rinfo->is_pushed_down)
+ if (only_pushed_down && !RINFO_IS_PUSHED_DOWN(rinfo, joinrel->relids))
continue;
if (rinfo->clause && IsA(rinfo->clause, Const))
OSDN Git Service
