Comment # 25 on bug 1111657 from Ludwig Nussel

AFAICT it's exactly what Matej referenced in comment#3. sin() gets called with
wrong prototype. libffi therefore does not pop the return value from the fpu
stack ie corrupting it. Short of fully reading and understanding the recently
added patch to glibc I can only guess that previously somehow the fpu
environment was reset or initialized between the faulty sin() and the actual
cos() call. IOW worked by coincidence. Pretty much x87 specific. The correct
fix is to fix the testsuite of python-cffi like upstream did.

Back to Matej.

Nevertheless to satisfy curiosity it would be interesting to know what glibc
did before to prevent this issue.