[J3] IEEE supported formats?

[Steven G. Kargl]

Thanks for the feedback, Malcolm.  Your reading of
the standard is how I read it.  The gfortran bug
report has identified other processors, that on
the same hardware, do not return the kind for an
Intel 80-bit type.  Those processors may not 
support that type under Section 17.

-- 
steve

On Fri, Jun 12, 2020 at 09:29:53AM +0900, Malcolm Cohen via J3 wrote:
> Hi Steve,
> 
> The standard literally says “an ISO/IEC/IEEE 60559:2011 floating-point format”. There is no qualification there, so that means, “an ISO/IEC/IEEE 60559:2011 floating-point format”.
> 
>  
> 
> There is absolutely no doubt that 60559:2011 defines several basic formats, also interchange formats (a superset), and extended formats (additional). There is no doubt that these are all floating-point formats.
> 
>  
> 
> Therefore, I see no bug in gfortran here. 80-bit “double-extended” is a perfectly valid format according to 60559:2011 subclause 3.7.
> 
>  
> 
> No, the standard should not say “including extended formats”, as that would implicitly exclude any interchange format that is neither basic nor extended, viz binary16. The unqualified “format” is correct.
> 
>  
> 
> Cheers,
> 
> -- 
> 
> ..............Malcolm Cohen, NAG Oxford/Tokyo.
> 
>  
> 
> From: J3 <j3-bounces at mailman.j3-fortran.org> On Behalf Of Steven G. Kargl via J3
> Sent: Friday, June 12, 2020 4:16 AM
> To: J3 Fortran <j3 at mailman.j3-fortran.org>
> Cc: Steven G. Kargl <kargl at troutmask.apl.washington.edu>
> Subject: [J3] IEEE supported formats?
> 
>  
> 
> A bug has been reported with gfortran's IEEE support, but it is
> unclear (to me) whether it is a compiler bug or a defect in the
> Fortran standard.
> 
> On some architectures, gfortran supports 4 REAL kinds with a radix
> of 2. For this discussion call these types REAL(4), REAL(8), REAL(10),
> and REAL(16). The types with kind=4, 8, and 16 map to IEEE Std 754-2008's
> binary32, binary64, and binary128 basic formats. REAL(10) maps to an
> extended binary64 format (sometimes called extended double or Intel
> 80-bit format). This program
> 
> use ieee_arithmetic
> print *, precision(1._4), selected_real_kind(6), ieee_selected_real_kind(6)
> print *, precision(1._8), selected_real_kind(8), ieee_selected_real_kind(8)
> print *, precision(1._10), selected_real_kind(16), ieee_selected_real_kind(16)
> print *, precision(1._16), selected_real_kind(21), ieee_selected_real_kind(21)
> end
> 
> outputs 
> 
> 6 4 4
> 15 8 8
> 18 10 10
> 33 16 16
> 
> It is claimed, that for the 3rd line, gfortran's ieee_selected_real_kind(16)
> should return 16. That is, it should select the binary128 basic format.
> 
> Is the intent of the Fortran standard to have Section 17 apply only to
> the IEEE Std 754-2008's basic formats? If yes, should the Fortran standard
> explicitly state this? If no, and extended formats can be supported by
> a processor, should the Fortran standard state that extended formats 
> are allowed. A search of the WD 1539-1 J3/18-007r1 did not turn up 
> any restrictions of the supported format. I'll note that it is possible
> that I missed the restrictions or in the absences of any restrictions
> gfortran result conforms to the Fortran standard.
> 
> -- 
> Steve
> 
> 
> 
> Disclaimer
> 
> The Numerical Algorithms Group Ltd is a company registered in England and Wales with company number 1249803. The registered office is: Wilkinson House, Jordan Hill Road, Oxford OX2 8DR, United Kingdom. Please see our Privacy Notice <https://www.nag.co.uk/content/privacy-notice>  for information on how we process personal data and for details of how to stop or limit communications from us.
> 
> This e-mail has been scanned for all viruses and malware, and may have been automatically archived by Mimecast Ltd, an innovator in Software as a Service (SaaS) for business.
> 

-- 
Steve