Failure of implicit real*8

[Jerry DeLisle]

On 02/19/2011 06:13 PM, Michael D. Berger wrote:
> Using:
>     gfortran44 -fdefault-real-8
>
> where:
>     [~]$ gfortran44 --version
>     GNU Fortran (GCC) 4.4.0 20090514 (Red Hat 4.4.0-6)
>
> I have two nearly identical subroutines very similar to:
>
>     subroutine orderRows(mx, idxSort, nrows, ncols)
>     dimension idxSort(nrows),mx(nrows,ncols),tmp(nrows,ncols)
>     ...
>
> where mx is expected to be real, in particular, real*8 because
> of the compile option.
>
> But one of the two interprets mx as an integer, giving
> radically wrong results.
>
> Using, instead:
>
>     subroutine orderRows(mx, idxSort, nrows, ncols)
>     dimension idxSort(nrows)
>     real mx(nrows,ncols),tmp(nrows,ncols)
>     ...
>
> produces correct results.
>
> I compiler bug?
>

First, not likely a compiler bug, though one can not tell without seeing more of 
the code, perhaps you can send a small working example that shows the problem.

Second, it is ill advised to ever use -fdefault-real-8.  Results are not always 
predictable.  Your code may not be as clean as you think and there could be side 
effects.  For example if there is a kind explicitly set somewhere.  Also, mx 
will be implicitly an integer if you do not explicitly define it to be real. So 
I think you were lucky that it worked at all before.

Third, if you need to change kind do this:

	real(kind=mp) mx(nrows,ncols),tmp(nrows,ncols)

and have mp be a parameter you set in one place ahead of where you use it.  This 
way you have explicit control over what is happening.  I would also recommend to 
always use IMPLICIT NONE so you don't have unknowns floating around in your code.

Fourth, consider getting a more recent version of the compiler. 4.5 is out now 
and 4.6 is approaching release (maybe a month).  There have been lots of bugs 
fixed, although 4.4 is OK.