This could be the same bug as #20580, but I am not really qualified to say, so I'm submitting a separate report. This bug came to light in the R language (see http://www.r-project.org and R-base on Red Hat Powertools), causing the random number generator for the Poisson distribution to misbehave. The bug manifests itself only when the code is optimized. I am attaching a small example program - abstracted from the R source code - that reproduces the bug. When compiled with gcc -lm, the output is 9 10 When compiled with gcc -lm -O2, the output is 9 nan If this works for you, a larger program is available that is much closer to the original source.
Created attachment 5789 [details] Example program to show bug
Actually, this is not related to #20580, but #21376 and seems to be fixed by that patch as well. So if the patch gets accepted into CVS gcc, it will be part of gcc-2.96-65 soon.