BTW, the gamma computation shouldn't be at all fragile, unless I've messed 
up somewhere:  the Lanczos sum is evaluated as a rational function in 
which both numerator and denominator have only positive coefficients, so 
there's little danger of nasty cancellations boosting the relative error.
I'd be surprised if use of fma instructions made more than 1 or 2 ulps 
difference to the result, but I'll take a look.