Forcing 64 bit “words” on a 32 bit machine should be relatively straightforward to implement, but I question the utility of this (factor of 2 space usage). Better would be to change the object representation so that the header of objects would be always 64 bits (or better, split into a 32 bit length and 32 bit type and GC flags). The implementation complexity of both is about the same (roughly 2 days of work).
However, too much work for me right now given the low need for this feature.
Aren’t 32 bit architectures pretty much a thing of the past? These days, even embedded systems are 64 bit architectures.