On 09/16/10 09:19, Paolo Bonzini wrote: > On 09/15/2010 09:31 PM, Jes Sorensen wrote: >> Floating point is just plain wrong. If someone wants to do something >> like in your example they really ask for an error. > > An error, not an overflow. > > Adding overflow checking on top of your patch is also fine. Another > possibility is to look ahead for the multiplier so that you correctly > base the divider and do everything in 64.64 fixed point. But it seems > overkill compared to floating-point, whose 53-bit mantissa precision > will almost always lead to exact results (large numbers usually have a > lot of zeros at the end, both in binary and in decimal).
I think it would be quite reasonable not to accept anything more than say 3-4 decimal points, since there are the t/g/m/k options as well. Cheers, Jes
