From:	Paul Koning 1695 <pkoning AT chipcom DOT com>
To:	djgpp mailing list <djgpp AT sun DOT soe DOT clarkson DOT edu>
Subject:	RE: deadly optimization
Date:	Mon, 16 Jan 95 14:14:00 PST
Encoding:	45 TEXT

>With 2.6.x there is no visible difference in speed between gcc and Watcom,
>which I think is the best optimizing compiler widely available for 386+.
>With such a beast any code handcrafting is a waste of time. Note how much 
time
>you spend writing inline code, and count how much time it can save during
>execution...

That's a generalization that may often be true but is not even close to true 
in
some cases.  High speed real time work may very well benefit substantially
from hand coding.  About a year ago I worked on a device that did packet
switching in 68040 software; I got somewhere between 1.5x and 2x performance
compared to C (from GCC) by careful hand-coding.  Note "careful".   If you 
don't
understand EVERY detail of instruction timing and cache behavior of your
particular platform, you don't know enough to do this.  However, if you do, 
and
you have skill, and you design the data structures right as well, then you 
stand
to gain a lot.

The other question is whether that work is worth it.  Sometimes it is; often 
it will
not be.  But if you think it is, do the study, don't take the opinion of 
people who
think that compilers are as good as the best programmers.  They aren't, not
by a long shot.

(Note that a 68040 is a simpler architecture than an x86 by a large margin, 
but
that doesn't really change the point.  For that matter, it's valid for RISC 
machines
too, popular "wisdom" notwithstanding.)

>Then I also hope we will not be stuck with CrazyGlue to x86 CPUs for the 
rest
>of our lives, so investing the precious time to study the architecture of 
the
>ies going to its decline is not very useful.

Amen.

     paul koning
     pkoning AT chipcom DOT com

- Raw text -