[Bug libfortran/35862] [F2003] Implement new rounding modes for run time

["burnus at gcc dot gnu dot org" <gcc-bugzilla at gcc dot gnu dot org>]

------- Comment #10 from burnus at gcc dot gnu dot org  2009-10-01 07:35 -------
(In reply to comment #9)
> Rounding modes are now implemented for formatted WRITE operations.  I do not
> have a clear enough idea of what the rounding modes really mean for READ
> operations.  For example:
> 
> "Rounding is done for all I/O, including input. For example, if you have:
>   read ("0.1", *, rounding="up") x
>   read ("0.1", *, rounding="down") y
> then you will have x == nearest(y, -1.)"

For REAL(8) I would expect for 0.1_8 with rounding up and down:
  0.10000000000000001
     9.99999999999999917E-002
or written as binary:
11111110111001100110011001100110011001100110011001100110011010 (= 0.1 rnd up)
11111110111001100110011001100110011001100110011001100110011001 (= 0.1 rnd down)

> Is this really what is meant by rounding up or down?
I think so.

> The internal representation of .1 may already be rounded up or down
> from a binary representation that carried more bits in the conversion
> process.  If we do not have access to more bits, how are we to know
> whether it is not already "half way in between"

Well, the rounding has to be based on the string (= decimal) number; if you
already have a binary number, it is too late.

> Similarly with nearest.

Using "nearest" with "-1." in this example only works if you know that "read"
automatically rounds up for 0.1.

 * * *

Handling this efficiently and reliably is presumably nontrivial.
See for instance:
  http://netlib.org/fp/ (esp. http://netlib.org/fp/gdtoa.tgz)
  http://dx.doi.org/10.1145/989393.989430
  http://www.informatik.uni-trier.de/Reports/TR-08-2004/rnc6_10_hack.pdf
and references therein.


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=35862