>  > One of the points I was trying to make is that, in certain cases I can
>  > document pretty well, these "micro optimizations" can do quite a bit.
> 
> You can state that all you want.  However, I'm not going to take it
> seriously until you can provide said documentation as to how it makes
> real life applications run faster.

One of Dave's optimizations made us go from 6.8 Mb per sec over
loopback go up to 7.4 Mb per sec on a 40Mhz MicroSparc.  I wonder, on
the TCP benchmark, how many megs are you moving per second on a
microsparc nowadays?  We have all those optimizations documented
together with the benchmark information, you can take a look at this
in the SparcLinux archives (search for lmbench results, they usually
come with an explanation on what was optimized).

Real optimizations for real life applications is an ftp transfer that
can fill a 100 Mbps pipe from a ss10.

Dave actually goes and measures the speed of his code before commiting
it to the tree.  He does not go and replaces a C routine with an
assembly routine that actually performs slower (sound familiar?  the
NetBSD checksum code had this very problem, I think Theo already
pointed this out to you guys).

Miguel.