I would rather not do this.   It optimizes for the uncommon case where all the objects are identical.  The common case is slightly worse off because the identity test is performed twice, once before the call to Py_RichCompareBool() and again inside it.  Also, the PR adds a little clutter which obscures the business logic.  

Another thought, micro-benchmarks on the identity tests require some extra care because they are super sensitive to branch prediction failures (See https://stackoverflow.com/questions/11227809 ).  A more realistic dataset would be:

  x = 12345
  data = [x] * 100 + list(range(500))
  random.shuffle(data)
  data.count(x)