RFR(XXS): 8161651: Logic in ConnectionGraph::split_unique_types() wrongly assumes node always have memory input

[Roland Westrelin]

http://cr.openjdk.java.net/~roland/8161651/webrev.00/

There's a small bug in ConnectionGraph::split_unique_types() where edge
MemNode::Memory of a node is being accessed even if it's not guaranteed
to exist. It's harmless so far because the access always falls withing
bounds but it's not guaranteed to stay that way with future code change.

Roland.