Compiling Crystal Space with SJLJ

March 11th, 2019 by Diana Coman

Miserable failures and unwanted problems make great blog fodder1 and so here I am, writing yet another set of compilation notes. This time it's all about how to ditch the non-standard ZCX run-time for its well-behaved and otherwise similarly fast ancestor, SJLJ.

My project to compile is fairly complex as it links dynamically with a bunch of libraries and it includes code in C, CPP and Ada all together. The first naive attempt at simply setting the PATH environment variable to point to the ZCX GNAT (with dynamic linking) and then recompiling the server code itself (i.e. without recompiling the libraries) failed at linking stage:

Error relocating ./euserver: _Unwind_SjLj_RaiseException: symbol not found
Error relocating ./euserver: _Unwind_SjLj_Unregister: symbol not found
Error relocating ./euserver: _Unwind_SjLj_Register: symbol not found
Error relocating ./euserver: __gxx_personality_sj0: symbol not found
Error relocating ./euserver: _Unwind_SjLj_Resume: symbol not found
Error relocating ./euserver: _Unwind_SjLj_ForcedUnwind: symbol not found

The good thing about those errors are that they confirm at least that the SjLj run-time is indeed to be used, rather than ZCX. So the correct intentions are there but the result is still miserable failure. A quick report in the forum yielded immediate help as Stan kindly provided the very useful observation that such errors at the final linking stage can only mean that it is still trying to link with a non-SJLJ gcc standard lib. Outrageous and unpermitted and how does it dare2!! A poke of the executable euserver by means of ldd revealed that indeed, it is apparently linking the ZCX and and all of that because of the pesky Crystal Space (CS) library! After a bit of initial doom when I thought I might need to recompile the whole bloody toolchain manually with --enable-sjlj-exceptions or such something, some rummaging around in the SJLJ-istic GNAT's directory revealed that it actually has already everything that's needed - if only they'd be linked instead of the ZCX-isting versions! So it came as a relief really that it's "only" CS3 that needs to be recompiled too4. Onwards!

A quick look through CS's makefile made it plenty clear that the naive approach of a simple .gpr file is not going to work: basically CS suffers deeply from ifdefism since it caters to all-and-everyone it can think of. So at the moment no option other than going at it with simply the PATH set properly to the SJLJ GNAT and then running: make clean; ./configure --without-java --without-perl --without-python --without-3ds --with-cal3d=$HOME/eulora/cal3d . And then all of a sudden, the configure reports that it couldn't find zlib although obviously, zlib did not run away anywhere. Moreover, a closer inspection of the full output of configure reveals that it *does* actually find zlib (so it's not a matter of "don't know where it is") but it reports it as unusable as it can't find apparently the ...headers. After various attempts at feeding it the path to the headers in various ways5, I got a bigger hammer and simply ran (making sure that it is PRECISELY the same version I had otherwise in the ZCX-istic toolchain):

curl -sL "" -O
tar -xzvf zlib-1.2.11.tar.gz
./configure --prefix=$HOME/eulora/zlib
make install

Back in the CS dir, I ran again:

jam distclean
./configure --with-z=$HOME/eulora/zlib --without-java --without-perl --without-python --without-3ds --with-cal3d=$HOME/eulora/cal3d
jam -aq libs plugins cs-config

The configure above worked fine and it reported that it did find zlib this time and it's usable and all is good. And then the jam went ahead and compiled all sorts, some of them even *with* #include < zlib.h > so obviously all was well, right? Except:

In file included from ./include/csutil/archive.h:36:0,
                 from cs/plugins/filesys/vfs/vfs.cpp:39:
./include/csutil/zip.h:32:18: fatal error: zlib.h: No such file or directory
compilation terminated

So VFS6 is brighter than all the rest and it can't find zlib. Moreover, a manual run of the very same compilation line with added -I$HOME/eulora/zlib/include completed absolutely fine and then the jam -q libs plugins cs-config went through without any trouble at all. The rest was simply waiting for the rather slow and bloated build of CS. So at this point, one would need to chase about in the config and jam files of CS and vfs where is the place to add the missing include flag. Some searches through the whole autoconf+configure+makefile+jamfile revelead that the ZLIB.CFLAGS and ZLIB.LFLAGS are set correctly for Jam to use so the problem seems to be simply that the vfs plugin is not aware it needs to use them. Nevertheless, I am not very keen on spending even more time essentially learning more on all those tools that are set to be discarded as at some point everything will have to move to GNAT and gprbuild anyway. So for now there it stands, ugly as it is: manually re-run that compilation line for vfs and be done with it7

Once CS was compiled with ZCX, the rest went perfectly fine, server and tests and everything else. Do remember to fully clean your previous compilation though before recompiling for changing from ZCX to SJLJ as otherwise all sorts of weird errors will appear. With gprbuild that's as simple and fast as running in your project's directory where the .gpr file is:

gprclean -r

And that's all! I happily have the server using SJLJ and so I can finally go up rather than down a few levels, up this pit of unexpected troubles and back to what I was actually doing - migrating both server and client to Eulora's new communication protocol so that there can finally be a text client too and everything one wants!

  1. Not that I am really lacking either problems or blog fodder but at this rate I'll end up with a compilation notes category all by itself. Other problems, failures and assorted troubles want a place in the sun too! 

  2. Especially since all the .gpr files of the server code specifically include the option --nostdlib to make sure the standard lib is not included and then a direct specification of the musl, ZCX rpath and dynamic-linker to use. 

  3. I recompiled also Cal3d since CS depends on Cal3d but that was simply a matter of cleaning properly the previous compilation, configure and recompile. Cal3d compiles even with a basic .gpr file so by comparison to CS it's absolutely wonderful. 

  4. In hindsight of course it has to be recompiled since CS contains all but the kitchen sink and specifically for this part of interest all sorts of threads and thread management "utilities" that pull in all sorts.  

  5. A run of ./configure --help said that one can use --with-z=path-to-zlib to directly tell configure where zlib is; very nice, except it still wasn't enough. 

  6. Virtual File System; inside a supposedly graphics engine; you have no idea. 

  7. CS is really to be compiled only once per system and that's it. If it becomes something I need to compile more often then it will have to be trimmed and ported to gprbuild anyway. 

Comments feed: RSS 2.0

Leave a Reply