Ok, this bug was fixed but not the actual bug I was after. But finaly, thanks to libgc's debug message, I got it. Actually, libgc overwrite pthread_create (and pthread_sigmask, and a few more) to know when a new thread is started. But this overloading is done "softly", ie. by a define in a C header file.
The program I'm working on was composed of many compilation unit, of which only a few include libguile (which then include gc.h). Some of these compilation unit were calling pthread_create without including libguile, and then when these threads were garbage collecting libgc crashed. So, in conclusion : always include libguile.h in the compilation units calling pthread functions. I would have prefered another solution, but I think we can live with that, _if_ it's documented somewhere. As a side note, I find libguile 1.9 slower than libguile 1.8.7. I suppose, being an unstable version, it's compiled by default with many debug options on. Is there an easy way to configure it for fast execution, for benchmarking reason ?
