On 2020-10-28 14:35 -0400, John Cowan wrote:
> I think they have to be implemented separately.  Experiments with Java on
> x86_64 show that the payload is sign-extended when converting from binary32
> to binary64 format, so that the unsigned value of the payload changes.

OK, I'll get to that soon.

The implementation + chibi FFI stub is in my repo,
https://github.com/Zipheir/srfi-208 .

> Currently the sign bit of the NaN is carried in the sign bit of the
> payload, but maybe that should be reconsidered too, in which case make-nan
> should take three arguments:  sign bit, quiet bit, payload, such that a
> negative payload is added to 2^52 (binary64) or 2^23 (binary32).  I believe
> that was Emmanuel's original design.

Yes, that's what the original pre-SRFI had.  I've implemented it with
the sign bit argument.

--
Wolfgang Corcoran-Mathe  <xxxxxx@sigwinch.xyz>

"Industry has surrounded people with artifacts whose inner workings
only specialists are allowed to understand." --Ivan Illich