Kiva AGG renderer breaks when compiled with gcc 4.7+

[jwiggins]

https://mail.enthought.com/pipermail/enthought-dev/2013-February/032139.html

[alvorithm]

• I could not fix this by compiling with gcc 4.6 (setting env variables LDSHARED to gcc-4.6 -pthread -shared -Wl,-O1 -Wl,-Bsymbolic-functions -Wl,-Bsymbolic-functions -Wl,-z,relro and CC to gcc-4.6 -pthread; telling distutils in distutils.cfg did not work, see here
• I could fix it by finding an older _agg.so shared object that I had lying around on my disk as advised here

This issue teams together with the problem that Kiva will be compiled to unusable if using libfreetype6[-dev] 2.4.10 (Ubuntu 12.10) or 11 (Ubuntu 13.04) to make it hard for users of newer distributions with virtualenvs to get started with Chaco. In Ubuntu 12.10 I had to fight hard my package system to downgrade to libfreetype6[-dev] 2.4.8 (from 12.04), which does work.

[heeres]

I could fix this on ubuntu 13.04 (64 bit) by compiling with gcc 4.6 (with libfreetype 2.4.10 plugged in), or alternatively by replacing _agg.so with a binary from enthought Canopy. Would be great if this could get fixed somehow!

[gurglespuge]

I am also having this issue, would be great if this package could be repaired. Using Ubuntu 13.04 32 bit, confirmed error on 64 bit.
I will try libfreetype6 2.4.8 ....

[pankajp]

I believe this issue is not about the freetype problem, but segfaults in enable tests (image_test_case.py) when enable is compiled with gcc >= 4.7 (i checked with 4.8.2). Please redirect if this is not the case.

The segfault can be eliminated by compiling without optimizations LDFLAGS=CXXFLAGS=CFLAGS="-O0" , or by compiling with clang (version 3.3) at any optimization level CXX=CC="clang" .

Tracing in gdb, the problem seems to be that gcc eliminated the code to assign the pixf pointer to m_pixf in https://github.com/enthought/enable/blob/master/kiva/agg/agg-24/include/agg_image_accessors.h#L36 , probably inlining it in some call and thinking it to be unused or presence of an undefined behavior.

I'm not sure whether it is likely to be an optimization opportunity for clang which may also break in later versions or a bug in gcc which will get fixed :)

[anthonyfk]

I've discovered something interesting about this that I'd like to share. Unfortunately I don't have time to investigate this further, but, if I specify the list of options gcc will enable at -O3 (gcc -Q --help=optimizers -O3 | grep enabled | awk '{print $1}') everything works without segfault. However, if I specify -O3 directly, it segfaults. Seems like a gcc bug to me.

[corranwebster]

The suspect image accessor functions above are part of the original agg codebase. It's interesting to note that the forked version of agg 2.4 at http://sourceforge.net/projects/agg/ (warning: on SourceForge) has removed the const nature of these constructors (see http://sourceforge.net/p/agg/svn/94/ for the appropriate change) to avoid problems with dangling references.

[jdeschenes]

This issue is now resolved with #288.

[jwiggins]

This should probably be closed.