subselect.c

来自「PostgreSQL7.4.6 for Linux」· C语言 代码 · 共 1,118 行 · 第 1/3 页

		result = (Node *) prm;
	}
	else if (node->parParam == NIL && slink->subLinkType == MULTIEXPR_SUBLINK)
	{
		List	   *exprs;

		/* Convert the lefthand exprs and oper OIDs into executable exprs */
		exprs = convert_sublink_opers(lefthand,
									  slink->operOids,
									  plan->targetlist,
									  0,
									  &node->paramIds);
		node->setParam = listCopy(node->paramIds);
		PlannerInitPlan = lappend(PlannerInitPlan, node);

		/*
		 * The executable expressions are returned to become part of the
		 * outer plan's expression tree; they are not kept in the initplan
		 * node.
		 */
		if (length(exprs) > 1)
			result = (Node *) (node->useOr ? make_orclause(exprs) :
							   make_andclause(exprs));
		else
			result = (Node *) lfirst(exprs);
	}
	else
	{
		List	   *args;

		/*
		 * We can't convert subplans of ALL_SUBLINK or ANY_SUBLINK types
		 * to initPlans, even when they are uncorrelated or undirect
		 * correlated, because we need to scan the output of the subplan
		 * for each outer tuple.  But if it's an IN (= ANY) test, we might
		 * be able to use a hashtable to avoid comparing all the tuples.
		 */
		if (subplan_is_hashable(slink, node))
			node->useHashTable = true;

		/*
		 * Otherwise, we have the option to tack a MATERIAL node onto the
		 * top of the subplan, to reduce the cost of reading it
		 * repeatedly.	This is pointless for a direct-correlated subplan,
		 * since we'd have to recompute its results each time anyway.  For
		 * uncorrelated/undirect correlated subplans, we add MATERIAL if
		 * the subplan's top plan node is anything more complicated than a
		 * plain sequential scan, and we do it even for seqscan if the
		 * qual appears selective enough to eliminate many tuples.
		 */
		else if (node->parParam == NIL)
		{
			bool		use_material;

			switch (nodeTag(plan))
			{
				case T_SeqScan:
					if (plan->initPlan)
						use_material = true;
					else
					{
						Selectivity qualsel;

						qualsel = clauselist_selectivity(subquery,
														 plan->qual,
														 0, JOIN_INNER);
						/* Is 10% selectivity a good threshold?? */
						use_material = qualsel < 0.10;
					}
					break;
				case T_Material:
				case T_FunctionScan:
				case T_Sort:

					/*
					 * Don't add another Material node if there's one
					 * already, nor if the top node is any other type that
					 * materializes its output anyway.
					 */
					use_material = false;
					break;
				default:
					use_material = true;
					break;
			}
			if (use_material)
				node->plan = plan = materialize_finished_plan(plan);
		}

		/* Convert the lefthand exprs and oper OIDs into executable exprs */
		node->exprs = convert_sublink_opers(lefthand,
											slink->operOids,
											plan->targetlist,
											0,
											&node->paramIds);

		/*
		 * Make node->args from parParam.
		 */
		args = NIL;
		foreach(lst, node->parParam)
		{
			PlannerParamItem *pitem = nth(lfirsti(lst), PlannerParamList);

			/*
			 * The Var or Aggref has already been adjusted to have the
			 * correct varlevelsup or agglevelsup.	We probably don't even
			 * need to copy it again, but be safe.
			 */
			args = lappend(args, copyObject(pitem->item));
		}
		node->args = args;

		result = (Node *) node;
	}

	return result;
}

/*
 * convert_sublink_opers: given a lefthand-expressions list and a list of
 * operator OIDs, build a list of actually executable expressions.	The
 * righthand sides of the expressions are Params or Vars representing the
 * results of the sub-select.
 *
 * If rtindex is 0, we build Params to represent the sub-select outputs.
 * The paramids of the Params created are returned in the *righthandIds list.
 *
 * If rtindex is not 0, we build Vars using that rtindex as varno.	Copies
 * of the Var nodes are returned in *righthandIds (this is a bit of a type
 * cheat, but we can get away with it).
 */
static List *
convert_sublink_opers(List *lefthand, List *operOids,
					  List *targetlist, int rtindex,
					  List **righthandIds)
{
	List	   *result = NIL;
	List	   *lst;

	*righthandIds = NIL;

	foreach(lst, operOids)
	{
		Oid			opid = lfirsto(lst);
		Node	   *leftop = lfirst(lefthand);
		TargetEntry *te = lfirst(targetlist);
		Node	   *rightop;
		Operator	tup;

		Assert(!te->resdom->resjunk);

		if (rtindex)
		{
			/* Make the Var node representing the subplan's result */
			rightop = (Node *) makeVar(rtindex,
									   te->resdom->resno,
									   te->resdom->restype,
									   te->resdom->restypmod,
									   0);
			/*
			 * Copy it for caller.  NB: we need a copy to avoid having
			 * doubly-linked substructure in the modified parse tree.
			 */
			*righthandIds = lappend(*righthandIds, copyObject(rightop));
		}
		else
		{
			/* Make the Param node representing the subplan's result */
			Param	   *prm;

			prm = generate_new_param(te->resdom->restype,
									 te->resdom->restypmod);
			/* Record its ID */
			*righthandIds = lappendi(*righthandIds, prm->paramid);
			rightop = (Node *) prm;
		}

		/* Look up the operator to pass to make_op_expr */
		tup = SearchSysCache(OPEROID,
							 ObjectIdGetDatum(opid),
							 0, 0, 0);
		if (!HeapTupleIsValid(tup))
			elog(ERROR, "cache lookup failed for operator %u", opid);

		/*
		 * Make the expression node.
		 *
		 * Note: we use make_op_expr in case runtime type conversion function
		 * calls must be inserted for this operator!  (But we are not
		 * expecting to have to resolve unknown Params, so it's okay to
		 * pass a null pstate.)
		 */
		result = lappend(result,
						 make_op_expr(NULL,
									  tup,
									  leftop,
									  rightop,
									  exprType(leftop),
									  te->resdom->restype));

		ReleaseSysCache(tup);

		lefthand = lnext(lefthand);
		targetlist = lnext(targetlist);
	}

	return result;
}

/*
 * subplan_is_hashable: decide whether we can implement a subplan by hashing
 *
 * Caution: the SubPlan node is not completely filled in yet.  We can rely
 * on its plan and parParam fields, however.
 */
static bool
subplan_is_hashable(SubLink *slink, SubPlan *node)
{
	double		subquery_size;
	List	   *opids;

	/*
	 * The sublink type must be "= ANY" --- that is, an IN operator. (We
	 * require the operator name to be unqualified, which may be overly
	 * paranoid, or may not be.)  XXX since we also check that the
	 * operators are hashable, the test on operator name may be redundant?
	 */
	if (slink->subLinkType != ANY_SUBLINK)
		return false;
	if (length(slink->operName) != 1 ||
		strcmp(strVal(lfirst(slink->operName)), "=") != 0)
		return false;

	/*
	 * The subplan must not have any direct correlation vars --- else we'd
	 * have to recompute its output each time, so that the hashtable
	 * wouldn't gain anything.
	 */
	if (node->parParam != NIL)
		return false;

	/*
	 * The estimated size of the subquery result must fit in SortMem. (XXX
	 * what about hashtable overhead?)
	 */
	subquery_size = node->plan->plan_rows *
		(MAXALIGN(node->plan->plan_width) + MAXALIGN(sizeof(HeapTupleData)));
	if (subquery_size > SortMem * 1024L)
		return false;

	/*
	 * The combining operators must be hashable, strict, and
	 * self-commutative. The need for hashability is obvious, since we
	 * want to use hashing. Without strictness, behavior in the presence
	 * of nulls is too unpredictable.  (We actually must assume even more
	 * than plain strictness, see nodeSubplan.c for details.)  And
	 * commutativity ensures that the left and right datatypes are the
	 * same; this allows us to assume that the combining operators are
	 * equality for the righthand datatype, so that they can be used to
	 * compare righthand tuples as well as comparing lefthand to righthand
	 * tuples.	(This last restriction could be relaxed by using two
	 * different sets of operators with the hash table, but there is no
	 * obvious usefulness to that at present.)
	 */
	foreach(opids, slink->operOids)
	{
		Oid			opid = lfirsto(opids);
		HeapTuple	tup;
		Form_pg_operator optup;

		tup = SearchSysCache(OPEROID,
							 ObjectIdGetDatum(opid),
							 0, 0, 0);
		if (!HeapTupleIsValid(tup))
			elog(ERROR, "cache lookup failed for operator %u", opid);
		optup = (Form_pg_operator) GETSTRUCT(tup);
		if (!optup->oprcanhash || optup->oprcom != opid ||
			!func_strict(optup->oprcode))
		{
			ReleaseSysCache(tup);
			return false;
		}
		ReleaseSysCache(tup);
	}
	return true;
}

/*
 * convert_IN_to_join: can we convert an IN SubLink to join style?
 *
 * The caller has found a SubLink at the top level of WHERE, but has not
 * checked the properties of the SubLink at all.  Decide whether it is
 * appropriate to process this SubLink in join style.  If not, return NULL.
 * If so, build the qual clause(s) to replace the SubLink, and return them.
 *
 * Side effects of a successful conversion include adding the SubLink's
 * subselect to the query's rangetable and adding an InClauseInfo node to
 * its in_info_list.
 */
Node *
convert_IN_to_join(Query *parse, SubLink *sublink)
{
	Query	   *subselect = (Query *) sublink->subselect;
	Relids		left_varnos;
	int			rtindex;
	RangeTblEntry *rte;
	RangeTblRef *rtr;
	InClauseInfo *ininfo;
	List	   *exprs;

	/*
	 * The sublink type must be "= ANY" --- that is, an IN operator. (We
	 * require the operator name to be unqualified, which may be overly
	 * paranoid, or may not be.)
	 */
	if (sublink->subLinkType != ANY_SUBLINK)
		return NULL;
	if (length(sublink->operName) != 1 ||
		strcmp(strVal(lfirst(sublink->operName)), "=") != 0)
		return NULL;

	/*
	 * The sub-select must not refer to any Vars of the parent query.
	 * (Vars of higher levels should be okay, though.)
	 */
	if (contain_vars_of_level((Node *) subselect, 1))
		return NULL;

	/*
	 * The left-hand expressions must contain some Vars of the current
	 * query, else it's not gonna be a join.
	 */
	left_varnos = pull_varnos((Node *) sublink->lefthand);
	if (bms_is_empty(left_varnos))
		return NULL;

	/*
	 * The left-hand expressions mustn't be volatile.  (Perhaps we should
	 * test the combining operators, too?  We'd only need to point the
	 * function directly at the sublink ...)
	 */
	if (contain_volatile_functions((Node *) sublink->lefthand))
		return NULL;

	/*
	 * Okay, pull up the sub-select into top range table and jointree.
	 *
	 * We rely here on the assumption that the outer query has no references
	 * to the inner (necessarily true, other than the Vars that we build
	 * below).	Therefore this is a lot easier than what
	 * pull_up_subqueries has to go through.
	 */
	rte = addRangeTableEntryForSubquery(NULL,
										subselect,
										makeAlias("IN_subquery", NIL),
										false);
	parse->rtable = lappend(parse->rtable, rte);
	rtindex = length(parse->rtable);
	rtr = makeNode(RangeTblRef);
	rtr->rtindex = rtindex;
	parse->jointree->fromlist = lappend(parse->jointree->fromlist, rtr);

	/*
	 * Now build the InClauseInfo node.
	 */
	ininfo = makeNode(InClauseInfo);
	ininfo->lefthand = left_varnos;
	ininfo->righthand = bms_make_singleton(rtindex);
	parse->in_info_list = lcons(ininfo, parse->in_info_list);

	/*
	 * Build the result qual expressions.  As a side effect,