I'm able to run both CityHash64 and CityHash128 perfectly fine on my 32-bit 
system. (Actually, CityHash128 has the same amount of data as the de-facto md5, 
which returns 128-bit.)

Can you provide a more detailed use-case (or sample code) of what you're trying 
to do that the current code won't accommodate?

Can I take the higher (or lower) 32-bits and use them as a hash value? Is this 
expected to be as fast as if we implemented a CityHash32 directly?

All the CityHash functions are tuned for 64-bit processors. That said, they 
will run (except for the new ones that use SSE4.2) in 32-bit code. They won't 
be very fast though. You may want to use Murmur or something else in 32-bit 
code.

As far as converting a "n-bit hash function" to a "m-bit hash function" where m 
< n, my understanding is that for any "good" hash function, you can simply 
discard n - m bits and keep m bits. For m= 32, all the CityHash variants should 
work fine with this method.

So, in short, it may make sense to create a CityHash32, but we're probably not 
going to do so.