Re: symbol catenation and montgomery

[Kevin Ryde]

Werner Koch <address@hidden> writes:
>    
> I'd like to hear about any performance problems.

I notice mpi/config.links uses i586 code on an i686.  I think you
might get a performance boost by using the i386 code instead.  The
i586 code is rather specific to pentium P5 and runs poorly on P6,
P-II, P-III.  By my reckoning the inner loops run as follows.

              i386    i586
              code    code

        p6     9.0    17.0    cycles/limb
        k7     6.0     7.5

On a k6, which gets classed as an i586 usually, the i586 code is the
better of the two (12 c/l for the i386 code, 11 c/l for i586), so
nothing needing to be done there.