Re: [RFD] optimized versions of openssl
Christoph Martin <firstname.lastname@example.org> writes:
> as you might know, there is only one binary package of libssl0.9.6 for
> each architecture. These packages are all build for the least capable
> processor. eg. for i386 there is no optimisation for pentiums
> etc. This has the benefit that it works on every i386 compatible
> processor but it is slow on processors where there could be a lot of
> The idea is to have a standard libssl0.9.6 package with no
> optimisation and some optional packages like libssl0.9.6-i686 or
> libssl0.9.6-k7 which can replace libssl0.9.6.
Like mplayer (non debian) has.
In general don't do that. On a case by case basis, if the package in
question uses a lot of cpu time, as cryptography does, and is greatly
speed up go ahead and make such packages. The speedup should be weigth
against the extra work and space needed to achieve it.
> Do we have a general opinion or policy on optimized library versions?
> Other libraries (like libgmp?) could have the same issues.
> One problem is, that the build process and the names and number of
> packages would be depending on the architecture and can not be done
> with a standard rules files.
Why? You have library-cpu_type in the control file and thats for one
arch only. So only on that arch it gets build. The rules files checks
the archs and only builds the relevant debs.
> Do we have modells for this?
> What do you think?
I would like a source based systems for a very few things. The deb
includes the source and depends on the Build-depends. The postinst
compiles the library and installs it.
But thats something taking a lot of effort to get working and keep
working. Nothing for the general package.