decimal8

[Michel Hack (1-914-784-7648) <hack@xxxxxxxxxxxxxx>]

Why not?  The DFP format parameters also accomodate w=2 with J=0.
This gives 1-digit precision with emax=3 and emin=-2, with positive
finite values ranging from 0.01 to 9000.  There are no subnormals,
so a better-balanced range might be from 0.001 to 900 -- although
one normal number (namely Nmin=1e-3) would not have an inverse.

I was wondering about that emin = 1 - emax rule.  This leads to a
fairly unbalanced range, as quite a few large numbers have subnormal
inverses.  The only reason to have this emin is to guarantee that
each normal number, including Nmin, has an inverse.  If we were
willing to accept overflow for 1/Nmin, that would be the ONLY case
in all rounding modes except away-from-zero, and with that exception,
the range would be balanced: each normal number would have a normal
inverse, if we had emin = -emax.  But I see where the problem is:
this would imply an odd exponent range.  Given that there are TWO
reserved exponents for binary, and none for decimal, this would
require reserving an additional exponent value (except in the case
of decimal16 with w=0 that I discussed earlier).  If we choose to
shift the exponent range by two (with emin = -1 - emax), there would
be many small normal numbers without an inverse -- not good.

Oh well...

Michel.
Sent: 2007-12-30 00:02:31 UTC