1. Use of groups as arguments in post and start. We argued that it is
easier to cache a group than a list. There is another performance
penalty that we did not discuss: the cost of checking that each process
in the argument group is also a process in the window group. When the
input argument is a list of ranks, the test is simply an "in range" test
for each rank. With a group member list, the test becomes more
expensive. Implementors should check on this.
2. New semantic section. Pls check it is what I advertise it to be:
correct and readable.