At the moment only named variables use registers, everything else uses the stack. A planned feature for the future is a peephole optimizer pass, which should be able to lower some named variable usage to the stack as well (for example, that screenshot I posted above would just be 1 (>L:XYZ,meters) instead of using a register).
All that being said, I haven’t noticed any slowdown from register usage (tested on i5-8400H and GTX1050Ti Max-Q at 1080p).