multilib fun - devel packages

[Florian La Roche]

On Mon, Dec 12, 2005 at 05:46:46AM -0500, Jakub Jelinek wrote:
> On Mon, Dec 12, 2005 at 11:42:20AM +0100, Florian La Roche wrote:
> > On Sat, Dec 10, 2005 at 12:32:26AM +0100, David Woodhouse wrote:
> > > On Fri, 2005-12-09 at 19:55 +0100, Matthias Saou wrote:
> > > > Let me ask a silly question here : Why do we want to encourage people to
> > > > build 32bit stuff on 64bit installs?
> > > 
> > > On PPC we use mostly 32-bit packages, because 64-bit is fairly pointless
> > > for most things. The choice of ppc32 vs. ppc64 isn't the same as the
> > > choice of ia32 vs. amd64, because in the ppc case the 32-bit option is
> > > still actually a sane architecture with a sane number of registers --
> > > whereas on amd64 you gain more by using the extra registers than you
> > > lose to the natural inefficiency of 64-bit code.
> > 
> > 
> > Even on ppc we kind of move slowly over to use more and more 64bit apps,
> > as multilib on power prefers 64bit applications over the 32bit ones.
> 
> But that's just rpm bug, it shouldn't always prefer 64bit rpms over 32bit
> ones, instead decide based on how it was configured.
> For x86_64, it should be obviously defined to prefer 64bit, I think
> on s390x similarly (there 64-bit code can do direct calls rather than
> go through the literal pool all the time etc.).  On ppc and sparc
> it should on the other side prefer 32bit rpms over 64bit.

This would depend on at least 32bit /sbin/ldconfig to also be able to
work with 64bit ELF files. If that is fixed somehow within
glibc*, then this change to rpm could be tried out.

regards,

Florian La Roche