Those before-after-patterns look like they do not preserve any inner nodes in the transformation. So the sub-DAG that's a match in the larger DAG would always only have one output (the %r node). All inner nodes of a match will never drive other nodes outside the match. Correct?
If a node in your before-pattern is not preseverved across optimization, then you can us the number of child nodes as additional label for that node in the match. For example for algorithms that backtrack more labels on nodes are good as it leads to earlier backtracking.
