Using the current settings write(*,*) x sometimes does not print all
possible digits of real(16) variables on S/380.
The result is that gfortran.dg/default_format_2.f90 FAILs. That test
case contains effectively:
x = 1.0_16
do i = 0, 200
write (s,*) x
read (s,*) y
if (y /= x) res = res + 1
x = nearest(x,-huge(x))
end do
On S/380 this fails 7 times with the current setting; for these abs(x-y)
= 9.6296497219361792652798897129246366E-0035.
The solution is to simply print one more digit by default; one
additionally has to enlarge then the size of one character string in the
test suite as it is otherwise to short. I don't know whether it solves
some of the other failures on REAL(16) systems such as Debian's SPARC;
presumably it won't help on PowerPCs where the glibc seems to have a bug
(if I read PR24685 correctly).
The following patch has been build on x86-64 by me and bootstrapped and
regtested on S/390 by Andreas Krebbel.
OK for the trunk?
Tobias
2008-01-17 Tobias Burnus <burnus@net-b.de>
* io/write.c (write_real): Increase default precision for REAL(16) by one.
2008-01-17 Tobias Burnus <burnus@net-b.de>
* gfortran.dg/large_real_kind_form_io_1.f90: Enlarge string for internal I/O.