Re: [PATCH 3/5] eglibc: Update 2.13 to avoid multilib conflicts

[Mark Hatle <mark.hatle@windriver.com>]

On 7/25/11 8:59 AM, Phil Blundell wrote:
> On Mon, 2011-07-25 at 14:47 +0100, Richard Purdie wrote:
>> --- /dev/null
>> +++ b/meta/recipes-core/eglibc/eglibc-2.13/arch-ia32.patch
>> @@ -0,0 +1,5309 @@
>> +Sync the i386 and x86_64 headers into one common IA32 set of headers.
>> +
>> +The goal is to ensure that any headers produced in a 32-bit or 64-bit build
>> +are not only functionally equivalent, but actually the same in order to avoid
>> +file conflicts.
>> +
>> +The only remaining conflict is the bits/syscall.h.  This is dynamically
>> +generated, and so far I've been unable to figure out how to get both
>> +i386 and x86_64 to generate the same file.  We'll need to handle this
>> +in the recipe itself.
>> +
>> +Signed-off-by: Mark Hatle <mark.hatle@windriver.com>
> 
> This patch is missing an Upstream-Status.  It's also rather large and
> intrusive which makes it hard to review sensibly and seems like it might
> be a maintenance headache in the future.  I wonder whether it would be
> better to just put the 32-bit and 64-bit headers for eglibc in separate
> subdirectories (say /usr/include/32/... and /usr/include/64/...) and not
> bother even trying to patch them to be the same.

(still catching up on email)

Someone else asked a similar question as well.  I opened a yoctoproject bug on
this as an attempt for people to understand what the problem is and how it's
being resolved (above and beyond the explanation above).

http://bugzilla.pokylinux.org/show_bug.cgi?id=1291

It doesn't have an upstream-status on it, because I'm not exactly sure what to
do with it yet.. I was still evaluating when I went on vacation.

Some of the patches are obvious IMHO.  It's simple things like the headers not
being the same, but the contents being identical -- or slight formatting
variations in the files.  This should be something I can send upstream.

Most of the remaining items the x86_64 version of the header is "correct", but
the i386 version of the header simply doesn't have knowledge that x86_64 even
exists.  Adding that knowledge was as simple as copying the x86_64 version on
top of the i386 version.  I don't know if upstream would permit this or not.

The few remaining ones usually had assembly optimizations on the i386 version
that were not in the x86_64 version.  I merged these in order for the
optimizations to be retained.  Again, I don't know if upstream would permit this
to either the i386 or x86_64 headers.

The syscall.h is generated (as noted above).  The issue is that it's generated
differently [and subsequently used differently when building eglibc].  So using
the oe_multilib_header helper was the only reasonable alternative I could see.

As for maintenance, I see this as not terribly complex once you understand what
and why the patches exist.  Except for the two? headers that I hand merged,
everything else is obvious once you run a diff between the stock i386 and x86_64
versions.  There is no intent for "original content" to be in this patch.

(For people wondering, this generally isn't a problem on power or mips because
there is a single architecture directory that builds for both 32-bit and 64-bit.
 IA32 doesn't do this.. they have a separate 32-bit [i386] and 64-bit [x86_64]
architecture directory.. if I were asked what the root cause of the bug was, I'd
immediately point to that.  There really should be a single IA32 architecture
that is capable of building both endians.)

--Mark

> p.
> 
> 
> 
> _______________________________________________
> Openembedded-core mailing list
> Openembedded-core@lists.openembedded.org
> http://lists.linuxtogo.org/cgi-bin/mailman/listinfo/openembedded-core