g++ 2.96 has crashed on me twice now.  One time it crashed when I hadn't
put in the header that gave a definition for a class I was using.
The other time it crashed on some code that used the GNU named return value
optimization, which is non-standard, but supposedly supported by g++.

gcc GNATS tracking numbers: c++/611  c++/613

The compiler should not crash.

I entered these bugs into the GNATS system for gcc, and I recieved an
eventual message on gcc-announce claiming the gcc-2.96 was considered to be
a pre-release version, and while C, and FORTRAN were likely to still work,
the C++ compiler probably had a number of issues.
I understand some eagerness to put in new versions of things.  Especially
since the new gcc seems to have much better per-processor optimizations for
the x86 family of chips (including AMDs offerings), but I expect releases
to be stable.  That is, after all, a big part of what I pay RedHat for.  If
didn't care about stability, I'd download and compile the latest and
greatest for myself.
I find the instability of the gcc-2.96 C++ frontend to be disturbing.
I think a reasonable fix would be to provide a set of back-patches to move
gcc back to 2.95.2, and patches to the distributed programs that use C++ so
that they are binary compatible with gcc 2.95.2.