Avoid unnecessary shm writes in Parallel Hash Join.

Currently, Parallel Hash Join cannot be used for full/right joins,
so there is no point in setting the match flag.  It turns out that
the cache coherence traffic generated by those writes slows down
large systems running many-core joins, so let's stop doing that.
In future, if we need to use match bits in parallel joins, we might
want to consider setting them only if not already set.

Back-patch to 11, where Parallel Hash Join arrived.

Reported-by: Deng, Gang
Discussion: https://postgr.es/m/0F44E799048C4849BAE4B91012DB910462E9897A%40SHSMSX103.ccr.corp.intel.com

src/backend/executor/nodeHashjoin.c

@@ -454,7 +454,26 @@ ExecHashJoinImpl(PlanState *pstate, bool parallel)
 				if (joinqual == NULL || ExecQual(joinqual, econtext))
 				{
 					node->hj_MatchedOuter = true;
-					HeapTupleHeaderSetMatch(HJTUPLE_MINTUPLE(node->hj_CurTuple));
+
+					if (parallel)
+					{
+						/*
+						 * Full/right outer joins are currently not supported
+						 * for parallel joins, so we don't need to set the
+						 * match bit.  Experiments show that it's worth
+						 * avoiding the shared memory traffic on large
+						 * systems.
+						 */
+						Assert(!HJ_FILL_INNER(node));
+					}
+					else
+					{
+						/*
+						 * This is really only needed if HJ_FILL_INNER(node),
+						 * but we'll avoid the branch and just set it always.
+						 */
+						HeapTupleHeaderSetMatch(HJTUPLE_MINTUPLE(node->hj_CurTuple));
+					}
 
 					/* In an antijoin, we never return a matched tuple */
 					if (node->js.jointype == JOIN_ANTI)