AltiVec assist for subnormal floats in vectors

AltiVec assist for subnormal floats in vectors

The old CPU in a macppc traps AltiVec instructions when they encounter
denormal or subnormal floats.  Emulate most of them.  They operate on
vectors of 4 single-precision floats.  The emulations either use
scalar operations (so vmaddfp becomes 4 of fmadds) or a formula (like
vrsqrtefp's 1 / sqrt(b) = 1 / sqrt(b * 2**126) * 2**63).

I am forgetting to emulate some instructions (at least vrfin, vrfiz,
vrfip, vrfim).  If I don't emulate it, it will still cause SIGFPE.

Mac OS never emulated these instructions, but set AltiVec's "non-Java"
NJ bit (which changes all subnormal floats to zero).  FreeBSD also
sets NJ; NetBSD does SIGFPE; Linux emulates them.  The POWER9 running
OpenBSD/powerpc64 does them in hardware (without trapping).

ok kettenis@ miod@