[libgfortran] I/O of 128 bits integers

[FX Coudert]

Hi all,

Attached patch is an attempt at handling I/O of 128bits integers, and 
any other types of integer in the future.

The libgfortran code relied on int64_t being the largest integer kind: 
many parameters were passed as int64_t before being cast into their 
final type. This is obviously wrong on some platforms, such as x86_64. 
So, I propose we #define GFC_INTEGER_LARGEST to the largest integer kind 
available, and use this instead of int64_t.

The question is: how do we know, when compiling the library, which is 
the largest available integer kind for the target?

To do some testing, I defined GFC_INTEGER_LARGEST to be __int128_t on 
x86_64 and int64_t everywhere else. With that, the patch bootstraps and 
is regtested on i686-linux and x86_64-linux. On the latter, I can now do:

$ uname -mo
x86_64 GNU/Linux
$ cat a.f90
   character(len=60) c
   integer(kind=16) i, j

   print *, bit_size(0_16)
   print *, huge(0_16)
   j = 2**(bit_size(0_16)-1) - 2**7
   write (c,*) j
   read (c,*) i
   if (i /= j) call abort
   write (c,'(I60)') j
   read (c,'(I60)') i
   if (i /= j) call abort
   end
$ gfc a.f90 && ./a.out
  128
  170141183460469231731687303715884105727
$


I am waiting for comments on the general idea, the coding-style and 
choice of name for GFC_INTEGER_LARGEST, as well as idea for a general 
way to determine this type at compile-time. When everything is settled, 
I will submit a proper patch (with ChangeLog entries and all).

Thanks,
FX