[Floyd_Pfeffer]

I am writing a program with Lahey/Fujitu Fortran 95. I have a subroutine which calls another subroutine. On the first call to this subroutine, it give a -NaN (not a number) answer. It only does this on the first time the subroutine is called. All subsquent calls return the correct answer. I have checked for a divide by zero which usually results in a -NaN but can't see it. Here is the weird part. When I put a screen print (Print *, var) in the subroutine, the problem goes away?? When I remove the print statement the problem comes back.
Has anyone ever had this kind of problem? What other things can cause a -NaN error? Any help would be appreciated.
Floyd Pfeffer

[praveen]

Some more things to check

Check for uninitialized variables
check for accessing an array outside its range
sqrt of a negative number

[AliE]

Originally Posted by praveen

Some more things to check

Check for uninitialized variables
check for accessing an array outside its range
sqrt of a negative number

I would add also divisions by zero

Edit: sorry just realized you have already excluded this option. Are you using implicit none statment to exclude badly declared variables?

[FMDenaro]

The first thing to do is to compile without any optimization and using the debug options, check bounds array and so on. The run-time error should also provide the infos about the line of your error. While NaN is a signal of a division by zero, it could be do to a wrong access to the memory location.

[AliE]

Filippo has provided a good answer as usual.

In your makefile add the following flags:
-oO
-fbounds-check

In this way the compiler avoids vectorization and it gives you an error if you are using an index out of bounds.

It looks to me an error that you can fix just commenting the code and running it line by line to find out your issue.

If you want to use a machinegun against a little ant, then valgrind is what you are looking for.

[Floyd_Pfeffer]

Taking the optimization off solved the problem. Thanks to all that helped
Floyd Pfeffer

[FMDenaro]

Originally Posted by Floyd_Pfeffer

Taking the optimization off solved the problem. Thanks to all that helped
Floyd Pfeffer

Without changing the error in some line in the code? Be careful, this is a signal of an error.

[AliE]

Originally Posted by FMDenaro

Without changing the error in some line in the code? Be careful, this is a signal of an error.

Yes, Filippo is again right. Using -o0 can just mask your problem, actually -03 flag is not so unsafe. Also note that there is a huge performance gap between -o0 and -o3 in term of speed (at least order 2) , so you should really continue with debugging!

Try to download "valgring" and run the code until that point (valgrind --leak-check=yes myprog ). You shuold be able to find out more about your error.
Otherwise use the brute force and comment all the routine that is giving you the NaN and sequentially uncomment the lines and the code to find out the one that is giving you problems. It is a time consuming option, but can be effective if your routine is not too long.

[praveen]

Yes, dont ignore such errors. Changing optimization level is not the solution.

Try compiling/running with other compilers like gfortran. Use flags -Wall and also -fbounds-check which was suggested before. Look at all warnings and convince yourself they are not errors.