> OK, I've now got some C code with a wrong answer.  I'll send-pr
> this later unless someone tells me it is obviously off the wall.

I see the same.  This particular problem is apparently caused by
a bug somewhere in __ieee754_atan2 which is found in the libm.a
object module e_atan2.o.

This was found by linking the program with object modules
extracted from a 1.2.1 libm.a on an otherwise 1.3_ALPHA i386
system.

Now *why* this happens I don't know, but it's now at least
narrowed down to 931 bytes of object code ;-)

- H=E5vard