In the attached patch there is mpz_bitcount, which is an optimized version 
of bit_length which accepts only mpz; in mpmath bitcount has almost always
mpz as argument, so this provides some speedup. Another speedup
comes optimizing input processing in mpmath_normalize.
In issue 94 in http://groups.google.com/group/mpmath-issues
there is a patch to libmpf.py to be used with this patch.
The speedup in runtests is 5% on my Athlon XP 2600.