Hi,

I am trying to figure out how IEEE_MAX_NUM_MAG (and IEEE_MIN_NUM_MAG, IEEE_MAX_NUM, IEEE_MIN_NUM) are supposed to behave when both arguments are zeros with different signs.

From what I read in the text of the 2018 standard, we are in the case where "the result is either X or Y (processor dependent)". On the other hand, the Fortran standard refers to the IEEE operations, but those do not have processor-dependent behavior. Moreover, they have now been deprecated in more recent standards.

Is there any interpretation on the topic? I am missing something? Can someone give me an idea what compilers give for IEEE_MAX_NUM_MAG(-0., 0.) and IEEE_MAX_NUM_MAG(0., -0.) and IEEE_MAX_NUM_MAG(-0., -0.)?

Thanks,
FX

On Tuesday, June 6, 2023 at 4:08:16 AM UTC-7, FX Coudert wrote:
> Hi,
>
> I am trying to figure out how IEEE_MAX_NUM_MAG (and IEEE_MIN_NUM_MAG, IEEE_MAX_NUM, IEEE_MIN_NUM) are supposed to behave when both arguments are zeros with different signs.
>
> From what I read in the text of the 2018 standard, we are in the case where "the result is either X or Y (processor dependent)". On the other hand, the Fortran standard refers to the IEEE operations, but those do not have processor-dependent behavior. Moreover, they have now been deprecated in more recent standards.
>
> Is there any interpretation on the topic? I am missing something? Can someone give me an idea what compilers give for IEEE_MAX_NUM_MAG(-0., 0.) and IEEE_MAX_NUM_MAG(0., -0.) and IEEE_MAX_NUM_MAG(-0., -0.)?
>
> Thanks,
> FX

Of all the various Fortran compilers to which I have access, only ifort/ifx support these four functions, and they return the second operand as their result when both operands are zero.

On 6/6/2023 7:08 AM, FX Coudert wrote:
> I am trying to figure out how IEEE_MAX_NUM_MAG (and IEEE_MIN_NUM_MAG, IEEE_MAX_NUM, IEEE_MIN_NUM) are supposed to behave when both arguments are zeros with different signs.
>
> From what I read in the text of the 2018 standard, we are in the case where "the result is either X or Y (processor dependent)". On the other hand, the Fortran standard refers to the IEEE operations, but those do not have processor-dependent behavior. Moreover, they have now been deprecated in more recent standards.

The upcoming Fortran 2023 revises the descriptions of these functions,
removing the processor-dependent aspects and updating the revision of
ISO/IEC 60559 used to the 2020 edition. See
https://j3-fortran.org/doc/year/21/21-122r1.txt and
https://j3-fortran.org/doc/year/23/23-007.pdf

--
Steve Lionel
ISO/IEC JTC1/SC22/WG5 (Fortran) Convenor
Retired Intel Fortran developer/support
Email: firstname at firstnamelastname dot com
Twitter: @DoctorFortran
LinkedIn: https://www.linkedin.com/in/stevelionel
Blog: https://stevelionel.com/drfortran
WG5: https://wg5-fortran.org

On Tuesday, June 6, 2023 at 7:22:05 AM UTC-7, Steve Lionel wrote:
> On 6/6/2023 7:08 AM, FX Coudert wrote:
> > I am trying to figure out how IEEE_MAX_NUM_MAG (and IEEE_MIN_NUM_MAG, IEEE_MAX_NUM, IEEE_MIN_NUM) are supposed to behave when both arguments are zeros with different signs.
> >
> > From what I read in the text of the 2018 standard, we are in the case where "the result is either X or Y (processor dependent)". On the other hand, the Fortran standard refers to the IEEE operations, but those do not have processor-dependent behavior. Moreover, they have now been deprecated in more recent standards.
> The upcoming Fortran 2023 revises the descriptions of these functions,
> removing the processor-dependent aspects and updating the revision of
> ISO/IEC 60559 used to the 2020 edition. See
> https://j3-fortran.org/doc/year/21/21-122r1.txt and
> https://j3-fortran.org/doc/year/23/23-007.pdf
>
> --
> Steve Lionel
> ISO/IEC JTC1/SC22/WG5 (Fortran) Convenor
> Retired Intel Fortran developer/support
> Email: firstname at firstnamelastname dot com
> Twitter: @DoctorFortran
> LinkedIn: https://www.linkedin.com/in/stevelionel
> Blog: https://stevelionel.com/drfortran
> WG5: https://wg5-fortran.org

Amusingly enough, this tweak in F'202X may drop the number of conforming compiler vendors from one down to zero.

The situation with signed zeroes and intrinsic MAX/MIN is also a portability problem; see https://github.com/klausler/fortran-wringer-tests/blob/main/signed-zero-max-min.f90

On 6/6/2023 10:41 AM, Peter Klausler US wrote:
> Amusingly enough, this tweak in F'202X may drop the number of conforming compiler vendors from one down to zero.

Thanks for noting that. I have created a test case and will send the
report to Intel. They could change the behavior now and still conform to
F2018, as well as F2023. Note that ifort/ifx needs you to specify
-assume minus0 for this case (without that, it does not conform to F2018.)
--
Steve Lionel
ISO/IEC JTC1/SC22/WG5 (Fortran) Convenor
Retired Intel Fortran developer/support
Email: firstname at firstnamelastname dot com
Twitter: @DoctorFortran
LinkedIn: https://www.linkedin.com/in/stevelionel
Blog: https://stevelionel.com/drfortran
WG5: https://wg5-fortran.org

> Of all the various Fortran compilers to which I have access, only ifort/ifx support these four functions, and they return the second operand as their result when both operands are zero.

Thank you for your comments.
As of today, the development version of GCC/gfortran has them too. The implementation is not, at this stage, conformant to the future F'202X.

FX