------- Comment #7 from kargl at gcc dot gnu dot org 2007-11-22 19:00 -------
(In reply to comment #6)
> Created an attachment (id=14609)
--> (http://gcc.gnu.org/bugzilla/attachment.cgi?id=14609&action=view) [edit]
> Patch (reverted Rev 117584) plus test case
>
> Initial implementation by reverting Rev. 117584 as suggested by Steve.
>
> Problems:
> - nearest(x,-1) for negative denormal x does not give a smaller number
> - returned values are not the smallest possible:
> 4.9406564584124654E-324 vs. 1.11253692925360069E-308
> 1.4012985E-45 vs. 5.87747175E-39
To fix this problem, I believe that you need to set emin to emin - prec + 1
Something like
+ mpfr_set_emin ((mp_exp_t) (gfc_real_kinds[sgn].min_exponent -
gfc_real_kinds[sgn].digit + 1));
It might be a -1.
IIRC, mpfr always stores a normalized significand, so you need to
play with emin to properly account for gradual underflow.
--
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=34192
