- Go back to the old method of computing the mantissa, but when post
processing, expand the mantissa to slightly larger than maxMantissa,
then in doRoundDown, if the result is not exact, subtract one.
Finally, let doNormalize figure out the rounding of the result.