[gmx-developers] integer overflow in force.c:calc_avcsix

[Gerald Mathias]

Dear all,

I just came across a bug in 'calc_avcsix'  in
'src/mdlib/force.c' (gmx-version 3.1.4).
The routine calculates the average C6 constant
for the long-range dispersion correction to energy and pressure.

The line 433

  csix /= (natoms*natoms);

causes an integer overflow  with a compiler that uses four byte integers 
(as most do)  for

  natoms > SQRT(2^31)  = 46,341   

as the product of integers 'natoms*natoms'
is calculated before the cast to double is performed.
Thus, natoms should be casted to double beforehand.

The effect might not be too obvious in inhomogeneous systems,
as 'natoms*natoms' can still hold a rather large (tough undefined) number.
However, in a simulation of a large water box I was running, I had a factor
of ten in the dispersion energy and  an increase of  6% of the density.

Best regards,

Gerald

-- 
Gerald Mathias
Lehrstuhl fuer BioMolekulare Optik (BMO)
Oettingenstrasse 67
D-80538 Muenchen

Tel.: +49-89-2180 9224
Fax.: +49-89-2180 9202