[lsb-discuss] Multi-lib standards

[Bruce Dubbs]

Dennis Gilmore wrote:

> It really depends on the architecture.  ppc64 sparc64 and most risc based 64 
> bit arches the only thing you gain by running 64 bit software is access to 
> greater than 4GiB of ram.  unlike x86_64 where you gain extra registers and 
> functionality that is only available via 64 bit mode.  you actually lose by 
> running a pure 64 bit system.  you end up with bigger binaries that take 
> longer to load and use more memory.  that is why they generally run 64 bit 
> kernels and 32 bit userland.

That is in general agreement with my previous post.  Do you have any 
data about how much bigger the binaries are in non-x86/x86-64 
architectures?  Except for some limited embedded applications, I 
wouldn't think that the 10% size penalty for x86_64 that I saw would be 
considered significant.

>> Because maintaining two sets of libraries where one will do is
>> non-trivial and expensive.  At a minimum, all changes have to tested twice.

> it is trivial and still needed.  people are still making 32 bit cpus in mass 
> quantities today.  at least how Fedora builds its packages,  we use the 32 bit 
> rpms for multilib on 64 bit arches. we build one set of 32 bit and one set of 
> 64 bit rpms.  I also tend to run my kvm virtual machines 32 bit even though 
> the host is 64 bit.  all the early atom cpus were 32 bit  the amd geode is 32 
> bit, via ships 32 bit x86 cpus.  other arches like arm are currently only 32 
> bit.   You really can not make a general assumption on this based solely on 
> x86/x86_64 

I see your point.

>> I'm not sure you are using the term deprecate in the same way I am.  To
>> me it means that it is discouraged and at some indefinite point in the
>> future it will be removed.  Using that definition, I see no reason not
>> to declare a mixed 32-bit and 64-bit system as deprecated right now.
> 
> on non x86/x86_64 arches its desirable. I suggest that we don't do that.  If 
> that statement is mad it should be made clear that its for particular arches 
> only.

Since I don't have any recent non-x86/x86_64 experience, I can't dispute 
your point, but I would like to see some data to demonstrate what is 
gained and lost by running on these other architectures.

The fundamental question here is whether the additional performance 
warrants the additional complexity.

   -- Bruce