Issue Description:

For integer types larger than a word, the currently fall back on using the generic _convert operation. But we can do significantly better for Float16, Float, and Double; instead of going to a full soft-float computation, we should pull out the high-order 16, 32, or 64 bits of the integer, or 1 into the low bit if any set bits were lost, and then do a word convert (or two words, in the case of Double on 32b systems).