Mac Forum / Programming / Mac Programming / September 2004

if you're looking to write fast code, you're going about it the wrong
way.  choosing to implement everything in assembler is like building a
ship out of toothpicks.  it can be done, but it's the wrong level of
abstraction, and it will take you so long that you'll never get
finished.

a much more useful strategy is to write the whole app, then profile it.
macosx has shark, a very nice profiler.  i figured out how to use it in
20 minutes.  it doesn't matter how familiar you are with a body of code,
you will guess wrong about where it's spending its time.  if the code is
cross-platform, it's probably spending its time in different places on
each platform.

say you profile your app, and it turns out that your first-try md5
implementation, written in sloppy C, only accounts for .5 percent of the
program's total runtime.  therefore, even if you optimized it to death,
so that its runtime is ZERO, the overall effect on the program is
completely unnoticeable.  all that work for nothing.

if you've never profiled a particular body of code before, chances are
excellent that there will be low-hanging fruit, and that you can make
amazing improvements with very little effort.  for example, the program
i recently used shark on.  it turns out that it was wasting 65 percent
of its time in the system idle loop.  65 percent!  i had been far too
cautious about putting the program to sleep for long-ish periods, afraid
i'd burn up too much cpu time and make the whole system sluggish.  all i
had to do was reduce the length of a single usleep() call to speed up
the program dramatically.

MD5 calculation would be o-small compared to file i/o, you will
gain little converting to asm or even altivec. For quickest
calculation interleave file i/o and MD5 calculation.