more on quad

[David G. Hough on validgh]

Earl Killian points out that 113 significant bits products are
not easy to build on 53 significant bit multipliers.  I was thinking
more of building them in conjunction with 64 significant bit multipliers
associated with 64x64->128 bit doubled-precision integer multiplication.

And I was incorrect in my analysis of complex multiplication with fused
multiply-add.  Although the simplest approach

	t = xr*yr
	r = t - xi*yi

can have large relative rounding error, for 2X the fpop's you could do instead

	t1 = xr*yr
	t2 = xr*yr - t1
	t3 = t1 - xi*yi
	r  = t3 + t2

and it looks to me like the worst case bound on roundoff would be about three
times larger than the bound on roundoff in a correctly-rounded result.
Of course both are susceptible to intermediate exponent spill.

One unsettling fact about fused multiply-add is that
a*b+c*d might be evaluated either as

	t = a*b
	r = t+c*d

or

	t = c*d
	r = a*b+t

and the rounded results may differ.  But for most programs that's probably no
more trouble than how extended precision has usually been implemented.