Hi,

> I wonder if we really need to use floatformat_to_double instead of
> floatformat_to_doublest for the split_half cases.  It seems like that
> relies unnecessarily on the host double format.  Do things work better
> if you use floatformat_to_doublest and DOUBLEST variables, instead
> of floatformat_to_double (likewise floatformat_from_doublest)?

The attached patch replaces the use of floatformat_to_double with
floatformat_to_doublest, fixing the problem.

There are other uses of floatformat_to_double in the function, by i'm
not sure if we should just replace every call of that function with the
"doublest" version (likewise with the floatformat_from_double ->
floatformat_from_doublest replacement).

Regards,
-- 
Luis Machado
Software Engineer 
IBM Linux Technology Center