Re: CPU specific/optimized Debian builds ?
On Thu, May 23, 2002 at 09:33:50AM -0700, Thomas Bushnell, BSG wrote:
> Michael Stone <mstone@debian.org> writes:
> > FYI, gcc has a long history of mediocre optimizations, and it
[snip]
>
> Do you have specific bug reports to file? GCC makes it very easy to
> turn on or off optimizations in the default set for various
> processors. So if some new 386-inspired optimization slows down the
> magic-chip port, it's easy for the magic-chip port to just not use
> that optimization.
>
> So I wonder if you have submitted the relevant bug reports, or you can
> give anything more specific than the vague statement above.
What bug reports? If you compare gcs to other compilers on the same
platform, gcc typically performs worse. gcs also has a goal of
portability that other compilers do not, so the difference in
optimization isn't a bug so much as a difference in focus. This has
nothing to do with tuning available optimizations to perform better on a
specific chip. This does, however, mean that the fact that a particular
processor introduced a particular improvement should not automatically
lead to the conclusion that gcs can take full advantage of that
improvement. The only way to know would be to benchmark the program with
different compiler options--which is exactly what I asked for.
Off the top of my head, one reference is
http://www.nersc.gov/research/FTG/pcp/performance.html
Take a look at relative performance between e.g., the compaq compiler on
alpha and the gnu compiler on alpha, or the portland compiler vs. the
gnu compiler on intel. gcs has historically been a solid, portable
compiler, but not a speed king. AFAIK, the gnu compiler people are aware
of benchmark comparisons between gnu and other compilers, so this isn't
a revelation. I'm sure you can find papers with some clever use of
google, or by checking some journals. Note that gcc 3 has performance
increases over previous versions, and the picture may change. I have not
seen comparative assesments between gcs 3 and other compilers, and would
be interested in the new numbers. Note that I am not *against* changing
debian's compilation standards, but I do think that it's premature to
suggest that before someone quantifies the benefits.
On Thu, May 23, 2002 at 09:34:44AM -0700, Thomas Bushnell, BSG wrote:
> Quite right. So can you give us some hard numbers about specific
> optimizations which were added, and slowed down "new fancy chips"?
> And also, while you're at it, explain why you didn't file the relevant
> bug reports too.
You're all about bug reports. These aren't bugs. A bug would be "2+2=8
on gnu compiler". "gnu compiler slower than compaq compiler on compaq
system" isn't a bug. But at any rate I think you're wilfully ignoring
what I said and the context in which I said it. I was not making a claim
about whether distributing a debian release with a specific set of
optimizations would affect performance. Specifically, I said that the
topic had come up before and that no one demonstrated significant gains.
I then said that no one had presented any new numbers since that time.
There have been people who *claimed* that their recompiling made a
difference, but there have been no formal, reproducable benchmarks
wherein people compared a basic debian distribution against one with
some or all parts compiled with a set of optimizations. I don't think
it's unreasonable to request such data when someone suggests that debian
should change its distribution policies. I also don't see why that
suggestion would so infuriate you.
--
Mike Stone
--
To UNSUBSCRIBE, email to debian-devel-request@lists.debian.org
with a subject of "unsubscribe". Trouble? Contact listmaster@lists.debian.org
Reply to: