The trick here is that these tests depend on c1 getting in front of
c2... and that cannot be assumed given random thread scheduling algorithms,
it must be enforced with barriers