Re: Are we losing users to Gentoo?

[Nathanael Nerode <neroden@twcny.rr.com>]

Patrice Fortier wrote:

> Le mer 31/03/2004 à 22:01, Matthias Urlichs a écrit :
>> Hi, Sean Harshbarger wrote:
>> 
>> > It would be nice to have support for optimized
>> > binaries. IMHO it would help Debian out a lot.
Little did you know, it already exists....
Debian doesn't have your-favorite-arbitrary-highly-specific-to-me optimized
binaries.  But it does allow for optimized binaries based on the specific
processor; look at /usr/lib/i686/cmov and all that.

>> As far as I recall, the performance gains from compiling everything with
>> your-favorite-optimizer-settings aren't _that_ great -- typically smaller
>> than what you'd gain by waiting a couple of months before buying your
>> computer.
> 
> I do agree with you for most of the stuff we use on our computers.
> 
> But I'd like to point out that it could be interesting for a couple of
> packages like libc and cypto libs (like libssl) to have a binary package
> (with generic i686 support for example, as 686 is now at least 7 years
> old).
Little did you know, Debian has many of these already.  :-)

> 
> libc contains math functions, and considering that the 386 didn't even
> had an arithmetic coprocessor,
Debian basically requires a 387 if you have a 386; or the kernel can
emulate, but then it's dog-slow.

> we should get a significant speed up for
> these functions. Plus every applications use it, so it would benefit
> everyone of them.
Package libc6-i686
Look in /lib/tls/i686/cmov/* once you've installed it....

> We have something similar with crypto libs as they make an heavy use
> of maths/crypto functions which could greatly benefit from a generic
> i686 binary package.
/usr/lib/i686/cmov/libcrypto.so.0.9.7
/usr/lib/i686/cmov/libssl.so.0.9.7
(and similar /usr/lib/i586 and /usr/lib/i486 variants)

> There are maybe a couple of other libs/applis which could be
> interesting.
The kernel (which already has lots of variants).
ogle-mmx
ogle-altivec
atlas3-*
atlas2-*

> As you can see, this is quite lite for debian maintainers: only
> 2 (or so) packages, and only with new compile options compared
> to the original package (i386).
> 
> If this first phase is proven to be successful, we can later think
> about some desktop applications which are known be greatly improved
> by this (gimp, mozilla come to mind).
Applications or libraries which are *known* to be greatly improved should
certainly have wishlist bugs filed for optimized versions; providing actual
statistics would help, because maintainers are rightly wary of the phony
optimization syndrome (where it 'seems' to make a big difference to various
users, but doesn't really).

Note also that the current default compilation options for Debian compile
for i486, and "tune" for i686.  (Tuning optimizes for a particular
processor in the ways which don't make it incompatible with older
processors.)  (The exceptional package is the kernel, for which there's a
real-i386 version which emulates the missing 486 opcodes.)  

This could, by itself, make a meaningful speed difference on the average
ix86 machine (though that isn't at all why it was done).  This is new for
sarge, so only packages which have been recompiled with GCC 3.3 will get
the benefits, but that's most of them.

-- 
Make sure your vote will count.
http://www.verifiedvoting.org/