No subject

[Fred Tydeman]

 Subject:   divide on RS/6000
 
 We do perform the division on the RS/6000 using reciprocals, just as
 described in the paper.  However, under the covers of the machine,
 the exponent is one bit longer, so that there are no denorms needed to
 represent IEEE denorms.  Until the "microcode" finishes the division,
 all IEEE denorms are held as internal, normalized numbers with full
 53 bit mantissas.  That's why the hardware gets it right, and your
 code, which suffers denorms, is off by a bit.  You would have to treat
 as a special case, division problems where the result or the reciprocal
 of the denominator could become denorms, and rescale these cases to
 avoid trouble.
    Best regards,
      Peter Markstein