[fortran] huge() vs IBM extended double

[François-Xavier Coudert]

:REVIEWMAIL:

>         PR fortran/24096
>         * trans-types.c (gfc_init_kinds): Use one less for max_exponent
>         of IBM extended double format.

Tested on ppc-darwin, this patch actually does what it is supposed to do :)
gfortran.dg/large_real_1.f90 is not failing any more.

OK for mainline, with the addition below. No need to apply to 4.0 as
we don't have library support for such kinds on that branch. Sorry for
answering so late, but I only recently had a powerpc computer where I
can properly test.

> +      if (fmt->pnan < fmt->p)
> +       /* This is an IBM extended double format made up of two IEEE
> +          doubles.

This test is also true for (and only for) mips_extended_double (see
gcc/real.c for details). From my reading of gcc/real.c, it's not a
problem, but I think it should be noted in the comment.

FX