Re: While we're at it

On Fri, 2 Apr 2010, Simon Casady wrote:

> I agree, a very good paper worth reading more than once.
> However I can't resist adding a couple of more stories, thoughts on
> the subject of optimization.
> There once was a processor from att that had strlen as a single
> opcode, one complex instruction.  

I386 can also do this with REPZ.

> It was convenient but it turned out to actually a little slower, in
> the real world, than test increment loop with simple instructions.
> (also a cics rics comparison on the same cpu !). The problem was
> interrupts.  It took too long to save all that state and restart the
> instruction and it had to be interruptible.

> Point being you have to profile to know whats faster.

Exactly.


> The other experience I had was with an early version of AIX.  Being
> curious and not wanting to do the work my boss wanted done I did some
> test with disk IO.  I tested clib stdio, read write system calls(page
> size), and mmap. To my surprise stdio won every time.  Apparently the
> people at IBM knew more than I did.  The point here is that if you are
> not intimately acquainted with all the features and quirks of a
> system, just keep it simple and of course profile.

Stdio probably buffers more efficiently.

		--Dean


-- 
Av8 Internet   Prepared to pay a premium for better service?
www.av8.net         faster, more reliable, better service
617 256 5494