Red Hat Bugzilla – Bug 21188
gcc over-optimizes volatile double
Last modified: 2007-04-18 12:29:58 EDT
gcc 2.96 ignores the volatile keyword when optimizing C code. This bug came to light in the R language (see http://www.r-project.org and the R-base RPM on powertools), which uses a routine called "machar" to determine the machine characteristics at runtime. As a result of the bug, machar reports double precision to be 2^-63 instead of the correct value 2^-52 on an i686. A search of the gcc mailing list archives shows that this bug was found independently by Tom Truscott and reported by him on 2 August, but without any follow-up. See http://gcc.gnu.org/ml/gcc-bugs/2000-08/msg00037.html He includes a short program which aborts when the bug is triggered. See also http://gcc.gnu.org/ml/gcc-bugs/2000-08/msg00038.html for an obviously related bug. The machar routine from the R language is too long to include here, but you can grab a copy from http://calvin.iarc.fr/gcc/machar.c This includes a main() function that prints the value of the double precision. Compare with and without optimization.
Created attachment 5622 [details] Example C program to highlight bug
I've posted a fix for this at http://gcc.gnu.org/ml/gcc-patches/2000-11/msg01227.html and am waiting now for feedback on it. If it is accepted, it will make it into next gcc rpms.
It was not accepted, but it will take quite some time before Richard Kenner writes a fix he wants to see, so I'll be putting this patch into gcc-2.96-65 and above.