]> granicus.if.org Git - postgresql/commitdiff
Refactor join_is_removable() to separate out distinctness-proving logic.
authorTom Lane <tgl@sss.pgh.pa.us>
Thu, 7 Apr 2016 17:11:30 +0000 (13:11 -0400)
committerTom Lane <tgl@sss.pgh.pa.us>
Thu, 7 Apr 2016 17:12:31 +0000 (13:12 -0400)
Extracted from pending unique-join patch, since this is a rather large
delta but it's simply moving code out into separately-accessible
subroutines.

I (tgl) did choose to add a bit more logic to rel_supports_distinctness,
so that it verifies that there's at least one potentially usable unique
index rather than just checking indexlist != NIL.  Otherwise there's
no functional change here.

David Rowley

src/backend/optimizer/plan/analyzejoins.c

index f7f5714e621f12cf0f57d82cc69d758a75c626ed..3d305eb9d884fbaeb3c1094729531c0fdf1b5bc2 100644 (file)
@@ -37,6 +37,9 @@ static bool join_is_removable(PlannerInfo *root, SpecialJoinInfo *sjinfo);
 static void remove_rel_from_query(PlannerInfo *root, int relid,
                                          Relids joinrelids);
 static List *remove_rel_from_joinlist(List *joinlist, int relid, int *nremoved);
+static bool rel_supports_distinctness(PlannerInfo *root, RelOptInfo *rel);
+static bool rel_is_distinct_for(PlannerInfo *root, RelOptInfo *rel,
+                                       List *clause_list);
 static Oid     distinct_col_search(int colno, List *colnos, List *opids);
 
 
@@ -152,7 +155,6 @@ join_is_removable(PlannerInfo *root, SpecialJoinInfo *sjinfo)
 {
        int                     innerrelid;
        RelOptInfo *innerrel;
-       Query      *subquery = NULL;
        Relids          joinrelids;
        List       *clause_list = NIL;
        ListCell   *l;
@@ -171,38 +173,13 @@ join_is_removable(PlannerInfo *root, SpecialJoinInfo *sjinfo)
 
        innerrel = find_base_rel(root, innerrelid);
 
-       if (innerrel->reloptkind != RELOPT_BASEREL)
-               return false;
-
        /*
         * Before we go to the effort of checking whether any innerrel variables
         * are needed above the join, make a quick check to eliminate cases in
         * which we will surely be unable to prove uniqueness of the innerrel.
         */
-       if (innerrel->rtekind == RTE_RELATION)
-       {
-               /*
-                * For a plain-relation innerrel, we only know how to prove uniqueness
-                * by reference to unique indexes.  If there are no indexes then
-                * there's certainly no unique indexes so there's no point in going
-                * further.
-                */
-               if (innerrel->indexlist == NIL)
-                       return false;
-       }
-       else if (innerrel->rtekind == RTE_SUBQUERY)
-       {
-               subquery = root->simple_rte_array[innerrelid]->subquery;
-
-               /*
-                * If the subquery has no qualities that support distinctness proofs
-                * then there's no point in going further.
-                */
-               if (!query_supports_distinctness(subquery))
-                       return false;
-       }
-       else
-               return false;                   /* unsupported rtekind */
+       if (!rel_supports_distinctness(root, innerrel))
+               return false;
 
        /* Compute the relid set for the join we are considering */
        joinrelids = bms_union(sjinfo->min_lefthand, sjinfo->min_righthand);
@@ -291,7 +268,8 @@ join_is_removable(PlannerInfo *root, SpecialJoinInfo *sjinfo)
                        continue;                       /* not mergejoinable */
 
                /*
-                * Check if clause has the form "outer op inner" or "inner op outer".
+                * Check if clause has the form "outer op inner" or "inner op outer",
+                * and if so mark which side is inner.
                 */
                if (!clause_sides_match_join(restrictinfo, sjinfo->min_lefthand,
                                                                         innerrel->relids))
@@ -302,65 +280,11 @@ join_is_removable(PlannerInfo *root, SpecialJoinInfo *sjinfo)
        }
 
        /*
-        * relation_has_unique_index_for automatically adds any usable restriction
-        * clauses for the innerrel, so we needn't do that here.  (XXX we are not
-        * considering restriction clauses for subqueries; is that worth doing?)
+        * Now that we have the relevant equality join clauses, try to prove the
+        * innerrel distinct.
         */
-
-       if (innerrel->rtekind == RTE_RELATION)
-       {
-               /* Now examine the indexes to see if we have a matching unique index */
-               if (relation_has_unique_index_for(root, innerrel, clause_list, NIL, NIL))
-                       return true;
-       }
-       else    /* innerrel->rtekind == RTE_SUBQUERY */
-       {
-               List       *colnos = NIL;
-               List       *opids = NIL;
-
-               /*
-                * Build the argument lists for query_is_distinct_for: a list of
-                * output column numbers that the query needs to be distinct over, and
-                * a list of equality operators that the output columns need to be
-                * distinct according to.
-                */
-               foreach(l, clause_list)
-               {
-                       RestrictInfo *rinfo = (RestrictInfo *) lfirst(l);
-                       Oid                     op;
-                       Var                *var;
-
-                       /*
-                        * Get the equality operator we need uniqueness according to.
-                        * (This might be a cross-type operator and thus not exactly the
-                        * same operator the subquery would consider; that's all right
-                        * since query_is_distinct_for can resolve such cases.)  The
-                        * mergejoinability test above should have selected only OpExprs.
-                        */
-                       Assert(IsA(rinfo->clause, OpExpr));
-                       op = ((OpExpr *) rinfo->clause)->opno;
-
-                       /* clause_sides_match_join identified the inner side for us */
-                       if (rinfo->outer_is_left)
-                               var = (Var *) get_rightop(rinfo->clause);
-                       else
-                               var = (Var *) get_leftop(rinfo->clause);
-
-                       /*
-                        * If inner side isn't a Var referencing a subquery output column,
-                        * this clause doesn't help us.
-                        */
-                       if (!var || !IsA(var, Var) ||
-                               var->varno != innerrelid || var->varlevelsup != 0)
-                               continue;
-
-                       colnos = lappend_int(colnos, var->varattno);
-                       opids = lappend_oid(opids, op);
-               }
-
-               if (query_is_distinct_for(subquery, colnos, opids))
-                       return true;
-       }
+       if (rel_is_distinct_for(root, innerrel, clause_list))
+               return true;
 
        /*
         * Some day it would be nice to check for other methods of establishing
@@ -561,6 +485,152 @@ remove_rel_from_joinlist(List *joinlist, int relid, int *nremoved)
 }
 
 
+/*
+ * rel_supports_distinctness
+ *             Could the relation possibly be proven distinct on some set of columns?
+ *
+ * This is effectively a pre-checking function for rel_is_distinct_for().
+ * It must return TRUE if rel_is_distinct_for() could possibly return TRUE
+ * with this rel, but it should not expend a lot of cycles.  The idea is
+ * that callers can avoid doing possibly-expensive processing to compute
+ * rel_is_distinct_for()'s argument lists if the call could not possibly
+ * succeed.
+ */
+static bool
+rel_supports_distinctness(PlannerInfo *root, RelOptInfo *rel)
+{
+       /* We only know about baserels ... */
+       if (rel->reloptkind != RELOPT_BASEREL)
+               return false;
+       if (rel->rtekind == RTE_RELATION)
+       {
+               /*
+                * For a plain relation, we only know how to prove uniqueness by
+                * reference to unique indexes.  Make sure there's at least one
+                * suitable unique index.  It must be immediately enforced, and if
+                * it's a partial index, it must match the query.  (Keep these
+                * conditions in sync with relation_has_unique_index_for!)
+                */
+               ListCell   *lc;
+
+               foreach(lc, rel->indexlist)
+               {
+                       IndexOptInfo *ind = (IndexOptInfo *) lfirst(lc);
+
+                       if (ind->unique && ind->immediate &&
+                               (ind->indpred == NIL || ind->predOK))
+                               return true;
+               }
+       }
+       else if (rel->rtekind == RTE_SUBQUERY)
+       {
+               Query      *subquery = root->simple_rte_array[rel->relid]->subquery;
+
+               /* Check if the subquery has any qualities that support distinctness */
+               if (query_supports_distinctness(subquery))
+                       return true;
+       }
+       /* We have no proof rules for any other rtekinds. */
+       return false;
+}
+
+/*
+ * rel_is_distinct_for
+ *             Does the relation return only distinct rows according to clause_list?
+ *
+ * clause_list is a list of join restriction clauses involving this rel and
+ * some other one.  Return true if no two rows emitted by this rel could
+ * possibly join to the same row of the other rel.
+ *
+ * The caller must have already determined that each condition is a
+ * mergejoinable equality with an expression in this relation on one side, and
+ * an expression not involving this relation on the other.  The transient
+ * outer_is_left flag is used to identify which side references this relation:
+ * left side if outer_is_left is false, right side if it is true.
+ *
+ * Note that the passed-in clause_list may be destructively modified!  This
+ * is OK for current uses, because the clause_list is built by the caller for
+ * the sole purpose of passing to this function.
+ */
+static bool
+rel_is_distinct_for(PlannerInfo *root, RelOptInfo *rel, List *clause_list)
+{
+       /*
+        * We could skip a couple of tests here if we assume all callers checked
+        * rel_supports_distinctness first, but it doesn't seem worth taking any
+        * risk for.
+        */
+       if (rel->reloptkind != RELOPT_BASEREL)
+               return false;
+       if (rel->rtekind == RTE_RELATION)
+       {
+               /*
+                * Examine the indexes to see if we have a matching unique index.
+                * relation_has_unique_index_for automatically adds any usable
+                * restriction clauses for the rel, so we needn't do that here.
+                */
+               if (relation_has_unique_index_for(root, rel, clause_list, NIL, NIL))
+                       return true;
+       }
+       else if (rel->rtekind == RTE_SUBQUERY)
+       {
+               Index           relid = rel->relid;
+               Query      *subquery = root->simple_rte_array[relid]->subquery;
+               List       *colnos = NIL;
+               List       *opids = NIL;
+               ListCell   *l;
+
+               /*
+                * Build the argument lists for query_is_distinct_for: a list of
+                * output column numbers that the query needs to be distinct over, and
+                * a list of equality operators that the output columns need to be
+                * distinct according to.
+                *
+                * (XXX we are not considering restriction clauses attached to the
+                * subquery; is that worth doing?)
+                */
+               foreach(l, clause_list)
+               {
+                       RestrictInfo *rinfo = (RestrictInfo *) lfirst(l);
+                       Oid                     op;
+                       Var                *var;
+
+                       /*
+                        * Get the equality operator we need uniqueness according to.
+                        * (This might be a cross-type operator and thus not exactly the
+                        * same operator the subquery would consider; that's all right
+                        * since query_is_distinct_for can resolve such cases.)  The
+                        * caller's mergejoinability test should have selected only
+                        * OpExprs.
+                        */
+                       Assert(IsA(rinfo->clause, OpExpr));
+                       op = ((OpExpr *) rinfo->clause)->opno;
+
+                       /* caller identified the inner side for us */
+                       if (rinfo->outer_is_left)
+                               var = (Var *) get_rightop(rinfo->clause);
+                       else
+                               var = (Var *) get_leftop(rinfo->clause);
+
+                       /*
+                        * If inner side isn't a Var referencing a subquery output column,
+                        * this clause doesn't help us.
+                        */
+                       if (!var || !IsA(var, Var) ||
+                               var->varno != relid || var->varlevelsup != 0)
+                               continue;
+
+                       colnos = lappend_int(colnos, var->varattno);
+                       opids = lappend_oid(opids, op);
+               }
+
+               if (query_is_distinct_for(subquery, colnos, opids))
+                       return true;
+       }
+       return false;
+}
+
+
 /*
  * query_supports_distinctness - could the query possibly be proven distinct
  *             on some set of output columns?