John,
From your benchmark I see that you didn't provide results for boost::multiprecision::fixed_int, but rather for boost::multiprecision::** backends::cpp_int_backend. I tried to update boost multiprecision branch, but boost/mutliprecision/backends directory appeared to be empty.
Many thanks for the use case - in point of fact it's proved pretty hard to
get a general purpose solution anywhere near your special-purpose code - simply because your code is so simple
I wouldn't say so. The only simple thing about my extended_int implementation is that it implements 3 basic operations: addition, subtraction and multiplication. But I don't see any problems on extending it to support division and other arithmetic operations. My point is that for fixed integer class it's not a good idea to represent negative numbers as two's complement. The main point of my benchmark was to compare fixed int implementations. And those should not have much performance difference in construction of temporary variables. I think what we are seeing in Andrii's use case may generalize across
quite a few different big number use cases.
Exactly, the main point of usage of fixed big integer class is that it doesn't require you to think about temporaries creation (except overflowing stack, but that's another topic). I would say that for math computations it is the best option in 90% of cases. It would be good for Boost to have both big int implementations: with stack allocated backend and heap allocated backend. I was able to contact Arseniy Kapoulkine, who's big int implementation is under /sandbox/SOC/2007/bigint. Arseniy agreed to work with Boost group to solve any problems that prevent the standardization. I will try to come up with benchmarks of the current implementation based on the Voronoi code. It would be very nice to combine that implementation with boost multiprecision. Cheers, Andrii Unsubscribe & other changes: http://lists.boost.org/mailman/listinfo.cgi/boost