Le samedi 05 mars 2005 à 22:24 -0500, Jody Hagins a écrit :
On Sun, 06 Mar 2005 01:33:50 +0100 Guillaume Melquiond <guillaume.melquiond@ens-lyon.fr> wrote:
I can't tell you what it would be for this particular implementation, but I can give you some results I got when using GMP some times ago. I hope everybody will agree that GMP is a high quality implementation of big integers, and hence the results are relevant.
Hmmm. Did you try the C++ wrappers, which are already implemented with expression templates?
The tests were purely C tests, so they weren't using the C++ wrappers. But I have also used GMP C++ wrappers in other occasions, and I know they are able to remove the last temporary before assignation. So the results in C++ would have been similar for single arithmetic operations: the ET addition would have been 6 times faster than the dumb addition, and the ET multiplication would have been 2 times faster than the dumb multiplication. The GMP C++ wrappers do not reuse the temporaries though, so the more complex expressions would benefit from C hand-writing. For example, in the expression "a*b + c*d + e*f", the temporary needed to computed "c*d" won't be reused later to compute "e*f. I'm not even sure such optimizations can be done by strictly using C++ meta-programming though. It seems to me some kind of language support would be required. But maybe I'm overly pessimistic and a complex ET implementation would be able to reuse temporaries. Best regards, Guillaume