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

[Steve Kargl <sgk at troutmask dot apl dot washington dot edu>]

On Sun, Jan 14, 2007 at 10:40:10PM +0100, Dominique Dhumieres wrote:
> > What does -fdump-tree-original show?
> > If 'x = cos(x)' eventually calls an OS supplied cos()
> > function, then contact Apple.
> 
> The dump is
> 
> MAIN__ ()
> {
>   real8 y;
>   real16 x;
> 
>   x = 1.0e+0;
>   y = 1.0e+0;
>   x = __builtin_cosl (x);
>   y = __builtin_cos (y);
>   {
>     _gfortran_st_write (&dt_parm.2);
>     {
>       real16 D.954;
>       D.954 = x - (real16) y;
>       _gfortran_transfer_real (&dt_parm.2, &D.954, 16);
>     }
>     _gfortran_st_write_done (&dt_parm.2);
>   }
> }
> 
> What is the source of __builtin_cosl?

gcc/builtins.* adds the information to the compiler.  Where the
evaluation is actually done, I do not know.  I, however, suspect
that __builtin_cosl (x) is doing something bad.  Can you print
out the values of x and y?

-- 
Steve