[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: Sign of NaN results in DVJ's tables



On Jul 21, 2006, at 12:48 PM, Michel Hack (1-914-784-7648) wrote:

The notation +/-x, +/-y or +/-xy is used, where "xy" means "either
the x or the y operand".  Does the "+/-" mean that the sign can be
arbitrary?  That's how I would interpret it, given that plain "x"
or "xy" could have been used to denote the actual operand.

Can you (Dave) confirm that this latitude was intentional?

Others:  Do you agree that the sign of an arithmetically-propagated
NaN need not be constrained?  Think of implementations where the sign
of Mult or Div is computed as a blind Xor of the operand signs, and
others where NaNs are propagated blindly.

Yes, there is quite a bit of latitude with the signs of NaNs, exactly
due to the considerations you raise (at least historically).


Actually, the question also applies to arithmetically-generated QNaNs,
which are also denoted as +/-qNaN (e.g. for +Inf + -Inf).

Generated qNaNs we say even less about.
The sign and payload are up to the language/system to sort out.

Months of long debate were invested here.

Jeff

754 | revision | FAQ | references | list archive