Re: [Bitkeeper-announce] 586 and earlier users?

On Mon, Sep 20, 2004 at 07:40:51AM -0700, Kevin P. Fleming wrote:
> Larry McVoy wrote:
> 
> >OK, let's put this to bed because I've gotten quite a few messages.  We 
> >will
> >not turn on 686 optimizations at this time.
> 
> I'm not at all opposed to you providing 686-optimized builds; you'd just 
> have to also provide non-optimized builds, and that would be a lot more 
> work for your build systems (even though I'm sure it's automated).

In this case, the practice is, typically, to segregate the performance
sensitive routines in a shared library, to build 2 such libraries
with different compiler options, and to dynamically link with the
right one at run time. 

Now does anyone use bk on a 386? At least the 486 and up processors
have the bswap instruction, which speeds up ntohl() and htonl()
although I suspect BK uses mostly byte streams in its protocol.

Finally, on many benchmarks with gcc, scheduling the instruction
stream for the decoders of the P6 family has more influence on
performancce than the use of 686 instructions (showing that the 
P6 core is often limited by its decoders). Such scheduling has
has little impact on 486/Pentium/Pentium MMX/K6 processors.
Athlons, for their part, decode almost anything you throw at 
them extremely well.

        Regards,
        Gabriel
_______________________________________________
Bitkeeper-users mailing list
Bitkeeper-users@xxxxxxxxxxxx
http://bitmover.com/mailman/listinfo/bitkeeper-users
To unsubscribe from this list, go to the above URL, follow instruction at the 
bottom of the web page.