Hm.  Looks like gcc tries to use the highest precision
available to do fpu computations and only converts to
the final precision at the end.  On SPARC v9 that's
128-bit floating point.  However, 128-bit is emulated
on all UltraSPARC CPUs.  It appears we have bugs in the
kernel FPU emulation code, which should be fixed.  But
gcc should not be generating that code, unless
`long double' is really used in the source, 'cause it will
be constantly trapping into the kernel for emulation, 
which will cause performance to suck.

Eduardo