Security question?

[Jim Gettys]

On Fri, 2004-11-05 at 13:51 +0000, Alan Cox wrote:
> On Iau, 2004-11-04 at 16:41, Daniel Stone wrote:
> > 
> > Function calls are not cheap.
> 
> That depends a lot on your platform and compile options. Even x86 calls
> are often pretty cheap with regparm enabled. In addition the vendors
> spent a lot of time optimising the CPU's for this so there are hardware
> predicted call/return stacks for example.
> 

My fundamental point is that systems have been changing, the performance
data we have old, the performance tools we now have have improved
greatly, and we'd be fools to expend very much effort *before* going and
doing some more serious analysis.

For example, some of the tuning of the system was last done in
MicroVAX-II days (literally), and I guarantee you that VAX procedure
calls were very slow (cost of at least 10 simple instructions).  For
example, at the time a single procedure call in the inner X graphics
loops (either client or server side) represented a 10% total performance
hit.

I'm willing to make a strong bet we're now much more sensitive to cache
behavior than # or speed of instructions, for example.

Let's make sure we fight the right battles...
				- Jim