Hi Thomas, Jerry, I added my thoughts to the PR.
We can use frexp to split mantissa and exponent and proceed from there. Apply some scale factor here and there to get to the same as Intel and NAG. Let's continue the discussion in the PR. Cheers, Harald > Gesendet: Mittwoch, 22. April 2026 um 08:08 > Von: "Thomas Koenig" <[email protected]> > An: "Jerry D" <[email protected]>, "Harald Anlauf" <[email protected]>, > gfortran <[email protected]> > Betreff: Re: [patch] PR93727 Fortran 2018: EX edit descriptor > > Hi Jerry, > > > One thought that hit me is that I can work the needed bit manipulations > > if I assume IEEE 754 format. I think we support platforms that are not > > IEEE 754. > > I think you could use the information that is used for calculating > DIGITS for precision support. > > For IEEE, you may have to take into account that the leading one is not > added if the biased exponent equals zero (zero and denorm numbers). > > Also, I am not sure that we do support anything besides IEEE format, > at least not for KIND=4 and KIND=8. For POWER, there is still the > legacy "double double" format for KIND=16, but that is not really > of prime importance, and IBM is also trying to get away from this, > hence the invisible KIND=17 hack. > > Best regards > > Thomas
