Question: Truncated floating point (bug?)

I have run into some queer behaviour of my Maple 9 software. I gave the following commands:

##############################################################
 with(numtheory):
 xx := invcfrac([[0],[5,1,5,1,4,1]]);
                                                                                                                  1/2
                                                                                                      103   4 1173
                                                                                              xx := - --- + ---------
                                                                                                      199      199

 evalf(xx,30);
  0.170835931237938034194790775404

 evalf(%,20);

  0.17083593123793803419   <---(A)

#########################################################

No problem, so far. But on the other hand when I wrote directly
#########################################################
 evalf(xx,20);
  0.17083593123793803418  <---(B)

#########################################################

The result (A) and (B) should be the same, but they aren't (the last digit of B is incorrect).
Does anybody know why? Does it happen also in  recent versions of Maple?