NT 3.5 and 4.0 on 4 (IIRC) and 8 processors machines required to be compiled from source on the target host, the shipping kernel only supporting 2 processors and cross-compilation not being an option. Fun times :)
IIRC you had to recompile the Kernel and the HAL. I don't remember the details but compiling on anything but the very machine targeted wasn't practical.
Why wouldn’t cross-compilation be an option?