People should use standard types like 'int' or even 'long' where size is not required, this provides more flexibility, I have seen some people using uintXX in some loops (you people know who you are), these things should be avoided.
Speed is not gained by 64-bit processing, but speed is gained by making code which doesn't need to be emulated on a 64-bit machine, if someone is wanting to run a 64-bit version of mineserver they can compile it them selves, when we reach a more public release I'm sure we can have package releases for both 64-bit and 32-bit.
Some things like map generation are done using libnoise, which internally used doubles, most FPUs are 80-bit FPUs regardless of what you through into them, however speed will sometimes differ depending on things like strictness settings if things where using floats over doubles.
Such things are minimal optimizations and should not be the first thought for anyone, I believe its more important that people focus on major flaws in speed with some peoples programming style (e.g. ftol everywhere)
If we can get a programming standard created to help eliminate and some peoples potentially slower coding this would be useful, at some time during the coming holidays I'll probably jump in and refactor some of the slower points.