For implementations that do not want to support full double arithmetic, it appears that Forth-2012 permits an implementation where a double uses fewer bits than twice the number of bits in a single. That is, I argue that as long as a single has at least 32 bits (the minimum required range of a double), an implementation-defined encoding of doubles where the only valid representations of a double are ( u 0 ) and ( n.negative -1 ) is possible (even if the testsuite needs more work before passi...
