The Poisson random number generator in R fails to generate correct
samples. The tail of the distribution is truncated. This appears to be
due to a bug in gcc 2.96 that I have submitted as #21423.
Just a quick question about this. If R-base is built without optimizations, you
don't have any problems? Even with the other modules?
R-base builds and passes "make check" without optimizations, but I think it will
be unacceptably slow. Users of scientific sotware normally want to drive their
machines pretty hard and expect high performance.
I'm distributing R (via CRAN: www.r-project.org) RPMS for Red Hat 7.0 built with
gcc 2.95.2, which works well. An alternative for you might be to use kgcc in
place of gcc, but retaining g77 2.96 for the parts written in Fortran. This
worked for me, but it has been reported that this leads to problems dynamically
loading R packages written in FORTRAN (e.g. akima).
Hmm. If it would slow it down the releasing unoptimized packages is not a good
solution. Thanks for pointing that out.
Unfortunately my hands are tied on this until gcc is fixed :/ If I use kgcc and
then g77 on the fortran stuff as you mentioned, it will break other things
whcih is also unacceptable. My only good option is to keep bugging Jakub until
I'm deferring this until gcc is fixed. Once gcc is fixed I will rebuild and
release an errata for R-base.
I am closing this since and errata was released for this.