Re: Bug#246319: libcln3: Segfaults in cln::I_to_digits when converting numbers to binary
Richard B. Kreckel writes:
> However, there is nothing wrong with that statement. It appears to be a
> compiler error introduced in GCC-3.0.0 and finally fixed in GCC-3.4.0.
> It is only triggered when the option -fno-exceptions is turned on when CLN
> is compiled. Without that option, everything works fine. I've just
> reproduced this bug with GCC releases 3.1.1, 3.2.3, 3.3.2 and 3.3.3, and
> also with 3.3.4-prerelease from CVS, all vanilla and bootstrapped locally.
thanks for tracking this down.
> Right now, I see four alternative solutions:
> 1) Upload a new Debian package where -fno-exceptions is not turned on.
If you change interfaces, you'll have at least change the package name
and/or soname and recompile dependent packages.
> 2) Fix CLN by playing dirty tricks. For instance, printing the variable
> erg's address in src/integer/output/cl_I_print.cc:30 seems to fix the
> problem. It should be possible to somehow read that address without
> any side-effect.
preferred solution, unless
> 3) Fix GCC on the 3.3 branch. It would probably take me at least a full
> day to distill a useful testcase, however. Likely more, if it turns
> out to be a Heisenbug. And if this leads to a fix in the GCC 3.3
> branch is more then doubtful, I guess.
you are able to extract a reduced testcase. yes.
> 4) Use Debian's gcc-3.4 , making that a requirment for depending
> packages, too.
Not really an option because of changes in the C++ ABI and libstdc++.
> Soooo, what are the chances for gcc-3.4 hitting sarge as the default
> compiler? In my experience, this is a much better compiler than 3.3.3, in
> all respects, including reliability. The whole 3.3 series seems to be
> having some problems. In another context, it has already been suggested
> by GCC developers that the 3.3 branch may have taken place too early,
> causing patches not beeing applied properly and so on .
IMO no chance as the default compiler. Remember the time needed for
the last C++ transition. Maybe as an option, if the runtime libraries
built by 3.3 and 3.4 are compatible.