Update of /cvsroot/monetdb/pathfinder/compiler/algebra/opt
In directory sc8-pr-cvs16.sourceforge.net:/tmp/cvs-serv5132/algebra/opt
Modified Files:
opt_join_pd.c
Log Message:
-- Fix bug in join pushdown optimization.
(If we want to push an equi-join through a union operator we have to
take the underlying renaming projections into account as otherwise
the naming convention thinks that both inputs to the union are identical.)
Index: opt_join_pd.c
===================================================================
RCS file: /cvsroot/monetdb/pathfinder/compiler/algebra/opt/opt_join_pd.c,v
retrieving revision 1.34
retrieving revision 1.35
diff -u -d -r1.34 -r1.35
--- opt_join_pd.c 10 Dec 2007 15:10:00 -0000 1.34
+++ opt_join_pd.c 13 Dec 2007 09:00:49 -0000 1.35
@@ -1245,15 +1245,99 @@
if (diff)
break;
- /* This rewrite is only correct if the union operator
- is implemented as a union all operation. */
- *p = *disjunion (eqjoin_unq (L(lp), rp, latt, ratt,
- p->sem.eqjoin_unq.res),
- eqjoin_unq (R(lp), rp, latt, ratt,
- p->sem.eqjoin_unq.res));
- modified = true;
- join_pushdown_worker (L(p), clean_up_list);
- join_pushdown_worker (R(p), clean_up_list);
+ /* stop if this consistency check fails */
+ if (p->sem.eqjoin_unq.res != latt &&
+ p->sem.eqjoin_unq.res != ratt)
+ break;
+
+ /* We have to look at the union operator in conjunction
+ with its renaming projection operators underneath as
+ we otherwise state that the input is the same -- the
+ cleanup phase only looks at the column names to decide
+ how to resolve conflicts. */
+ if (L(lp)->kind == la_project && R(lp)->kind == la_project) {
+ PFalg_att_t latt1 = att_NULL,
+ latt2 = att_NULL,
+ res1,
+ res2;
+ PFalg_proj_t *proj_list1,
+ *proj_list2;
+ PFalg_att_t cur;
+ unsigned int count = 0;
+
+ /* create projection list */
+ proj_list1 = PFmalloc (p->schema.count *
+ sizeof (*(proj_list1)));
+ proj_list2 = PFmalloc (p->schema.count *
+ sizeof (*(proj_list2)));
+
+ /* Copy the projection list except for the join column. */
+ for (unsigned int i = 0; i < lp->schema.count; i++) {
+ cur = lp->schema.items[i].name;
+ /* consistency check to ensure that the projections
+ and the union operator are aligned */
+ assert (cur == L(lp)->sem.proj.items[i].new &&
+ cur == R(lp)->sem.proj.items[i].new);
+
+ if (cur == latt) {
+ /* remember old join names */
+ latt1 = L(lp)->sem.proj.items[i].old;
+ latt2 = R(lp)->sem.proj.items[i].old;
+ }
+ else {
+ proj_list1[count] = L(lp)->sem.proj.items[i];
+ proj_list2[count] = R(lp)->sem.proj.items[i];
+ count++;
+ }
+ }
+
+ /* Extend the projection list with the columns of
+ the right join input (again discarding the join
+ column. */
+ for (unsigned int i = 0; i < rp->schema.count; i++) {
+ cur = rp->schema.items[i].name;
+ if (cur != ratt) {
+ proj_list1[count] = proj (cur, cur);
+ proj_list2[count] = proj (cur, cur);
+ count++;
+ }
+ /* Make sure that we do not write more
+ projection than we allocated space for it. */
+ if (count == p->schema.count)
+ break;
+ }
+
+ /* Skip rewrite if we have duplicate columns. */
+ if (count != p->schema.count - 1)
+ break;
+
+ assert (latt1);
+ assert (latt2);
+
+ /* find out the correct join result columns */
+ res1 = latt1 < ratt ? latt1 : ratt;
+ res2 = latt2 < ratt ? latt2 : ratt;
+
+ /* fix the output schema by adjusting the result name */
+ proj_list1[count] = proj (p->sem.eqjoin_unq.res, res1);
+ proj_list2[count] = proj (p->sem.eqjoin_unq.res, res2);
+ count++;
+
+ /* This rewrite is only correct if the union operator
+ is implemented as a union all operation. */
+ *p = *disjunion (PFla_project_ (
+ eqjoin_unq (LL(lp), rp, latt1, ratt,
+ res1),
+ count, proj_list1),
+ PFla_project_ (
+ eqjoin_unq (RL(lp), rp, latt2, ratt,
+ res2),
+ count, proj_list2));
+ modified = true;
+
+ join_pushdown_worker (LL(p), clean_up_list);
+ join_pushdown_worker (RL(p), clean_up_list);
+ }
} break;
case la_fun_1to1:
-------------------------------------------------------------------------
SF.Net email is sponsored by:
Check out the new SourceForge.net Marketplace.
It's the best place to buy or sell services
for just about anything Open Source.
http://ad.doubleclick.net/clk;164216239;13503038;w?http://sf.net/marketplace
_______________________________________________
Monetdb-pf-checkins mailing list
[email protected]
https://lists.sourceforge.net/lists/listinfo/monetdb-pf-checkins
