Jeremy Kerr <jk@ozlabs.org> writes:
> Thanks for the benchmark app, thought I'd pitch in with some ppc
> results:
It looks to me like we should go with the lookup table approach, as
being the best tradeoff of speed improvement vs platform and compiler
independence. The "float hack" is right out ;-)
I wonder whether it is worth fooling with alternatives for the data type
of the lookup table entries. unsigned char might be a tad faster
(avoid useless sign-extension work). int might be faster yet, but it
would enlarge the table, creating a distributed overhead that the
microbenchmark would completely fail to show. Or we could buy that
back by reducing the table to cover only 6 bits, knowing that 12 is
more than we need.
regards, tom lane