This is the mail archive of the fortran@gcc.gnu.org mailing list for the GNU Fortran project.


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]
Other format: [Raw text]

Re: Problem with zeros.


Tim Prince wrote:
Tobias Burnus wrote:
Hi,

Éric Depagne wrote:
My code reads a file, and puts the content of this file into an array.
The array is defined like this:

DIMENSION X(40,30)
The content of the file looks like this:
9.999e-46 9.999e-46 9.999e-46 9.999e-46 9.999e-46 2.944E-19
Which gives the run-time error:

At line 2 of file xxx.f90 (unit = 5, file = 'stdin')
Fortran runtime error: Range error during floating point read

I would like to know if there is a way to let gfortran know that when he encounters values that are too small, he shall convert them to zero ?
To my knowledge not.

I played around with other compiles and:

a) 9.999e-46


1. ifort, NAG f95, sunf95, pgf95: denormalized number: 1.4012985E-45 2. g95, openf95: 0.0 3. gfortran: Range error

(gcc+glibc with scanf: (2).)

b) 1e99

1. sunf95, g95, f95, openf95, pgf95: Inf
2. ifort, gfortran: Range error

(gcc+glibc with scanf: (1).)

For (a) I think (1) is the best result and (2) is also ok. I don't see
any good reason to do (3).
For (b) one can argue whether (1) or (2) is better; I like (1) more, one
might argue that (2) has also its merits and one may argue for an
option, but overflow FP traps would have the same result, I'd guess.
(Assuming that the data is used later on.)

What do the others think?

At the time when we persuaded the Intel compiler team to institute the present behavior of ifort, the primary models for compatibility were g77 on linux, and CVF on Windows. Both of those compilers silently convert subnormal inputs to the nearest available sub-normal (0. if appropriate). Throwing an exception which terminates the program, in the absence of iostat handling, breaks important applications. So it was accepted that the historical behavior of g77 and CVF should be continued.

Thanks for the information. I am looking into why we do not get the sub-normals now.

Jerry


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]