Possible explaination why large_real_kind_form_io_2.f90 is failing on PPC/OSX

[Tim Prince]

Dominique Dhumieres wrote:
> Trying to understand why large_real_kind_form_io_2.f90 is failing on PPC/OSX,
> I ended up with the following code:
>
> ! { dg-do run }
> ! { dg-require-effective-target fortran_large_real }
> ! PR libfortran/24685
> program large_real_kind_form_io_2
>   ! This should be 10 or 16 on systems that support kind=10 or kind=16
>   integer, parameter :: k = selected_real_kind (precision (0.0_8) + 1)
>
>   print *, k
>   print *, nearest(1.0_8,1.0_8)
>   print *, nearest(1.0_8,-1.0_8)
>   print '(1pg50.40))', nearest(1.0_8,1.0_8)
>   print '(1pg50.40)', nearest(1.0_8,-1.0_8)
>   print *, nearest(1.0_k,1.0_k), nearest(1.0_k,1.0_k)-1.0_k
>   print *, nearest(1.0_k,-1.0_k), nearest(1.0_k,-1.0_k)-1.0_k
>   print '(2(1pg60.50))', nearest(1.0_k,1.0_k), nearest(1.0_k,1.0_k)-1.0_k
>   print '(2(1pg60.50))', nearest(1.0_k,-1.0_k), nearest(1.0_k,-1.0_k)-1.0_k
>   
> end program large_real_kind_form_io_2
>
> which gives:
>
>           16
>    1.00000000000000     
>      1.000000000000000222044604925031308084726    
>     0.9999999999999998889776975374843459576368    
>   1.0000000000000000000000000000000000        2.4651903288156618919116517665087070E-0032
>   1.0000000000000000000000000000000000       -1.2325951644078309459558258832543535E-0032
>     1.00000000000000000000000000000000000000000000000000        2.46519032881566189191165176650870696772900000000000E-32
>     1.00000000000000000000000000000000000000000000000000       -1.23259516440783094595582588325435348386400000000000E-32
>
> with two results that might explain the failure:
>
> (1) the default format is too short to distinguish close values, 
> so codes such that:
>
>   write (tmp, *) b
>   read (tmp, *) a, c
>
> may fail to read a and c as b(1) and b(2) if the default format gives
> ambiguous output.
>
> (2) The general format for real(16) seems to stop the decoding too
> soon (without reaching the last significant digits).
>
>   
The 33 digits shown in print* default  are sufficient to distinguish 
neighboring values for 128-bit IEEE format.  IEE754 specifically permits 
longer specified formats to display zeros for trailing digits beyond 
that point.