On Mon, Dec 11, 2006 at 12:15:51PM -0500, Ron wrote:
I'd say the fairest attitude is to do everything we can to support
having the proper experiments done w/o presuming the results.
Who's presuming results?[1] It is fair to say that making extraordinary
claims without any evidence should be discouraged. It's also fair to say
that if there are specific things that need cpu-specific tuning they'll
be fairly limited critical areas (e.g., locks) which would probably be
better implemented with a hand-tuned code and runtime cpu detection than
by magical mystical compiler invocations.
Mike Stone
[1] I will say that I have never seen a realistic benchmark of general
code where the compiler flags made a statistically significant
difference in the runtime. There are some particularly cpu-intensive
codes, like some science simulations or encoding routines where they
matter, but that's not the norm--and many of those algorithms already
have hand-tuned versions which will outperform autogenerated code. You'd
think that with all the talk that the users of certain OS's generate
about CFLAG settings, there'd be some well-published numbers backing up
the hype. At any rate if there were numbers to back the claim then I
think they could certainly be considered without prejudice.
