Comment # 4 on bug 1127564 from Richard Biener

So upstream decided this is a glibc issue which I filed here:

https://sourceware.org/bugzilla/show_bug.cgi?id=24346

you still need to use -fexcess-precision=standard for sphinxbase and work
around the glibc issue with forcing the results of libm routines not
specified by IEEE to double with for example

 volatile double tem = log (...);
 ... use tem ...

-O0 optimization should not be necessary.

That said, not sure whether i586 with i387 math is still relevant - there
are no longer any 32bit x86 CPUs in production without SSE2 support
(but the i586 distro didn't yet arrive here, besides for 32bit multilib
support in SLE15+ which requires a x86_64 CPU even for i586 built packages).

Re-assigning to Andreas.