🚀 go-pugleaf

In article <9209121817.AA13759@SPUNKY.CS.NYU.EDU> figueroa@SPUNKY.CS.NYU.EDU (Samuel A. Figueroa) writes:
>According to a paper I have by W. Kahan (the main person behind both IEEE
>Standards for Floating-Point Arithmetic, as well as PARANOIA; this paper
>complains about the quality of Cray's floating-point arithmetic), the IIT
>math coprocessor doesn't round extended (80-bit) precision correctly,
>though due to a "fluke" in their design, this problem doesn't show up in
>double (64-bit) precision.  Does PARANOIA complain about this?  (I would
>not be surprised if it didn't, since few compilers provide direct access
>to extended precision, so PARANOIA might not be able to test this.  How-
>ever, I haven't looked at PARANOIA's source code, so I don't know if it
>somehow knows how to figure this out.)

It does not, and I don't see how it could in a Fortran program where
the relevant variables are specified as double precision.  (Once the
variable is moved out of the fpu, it gets truncated anyway.)

Furthermore, in what context could this be relevant if double precision
is adequate?  (And if it were not, how could you ensure, in Fortran,
that you are using the full 80 bits precision where you need it?)

Luc B