On Tue, Nov 6, 2012 at 2:10 PM, Phillip Moore <[email protected]> wrote:
> On Tue, Nov 6, 2012 at 12:53 PM, Steven Jenkins
> <[email protected]> wrote:
>> I've been working on gcc recently (gcc 4.7.2, primarily on RHEL 5 &
>> RHEL 6) and am encouraged to see others are doing so (Note: I've not
>> been using the efs-deploy-config-gnu-gcc infrastructure, but I'm
>> looking at it as a baseline).  I have a couple of observations:
>
...
>> 2- Are there some test cases out there for multilib handling?  I've
>> put together some tests I'm doing for that handling, but it would be
>> great if we had a consistent test suite.  Once I have things converted
>> to this infrastructure, I'm happy to submit what I have.
>
> If by test case you mean a distribution that is built using the
> multilib functionality, I don't know.   First of all, the "test suite"
> for builds of gcc isn't even used, since it requires autogen, and I've
> never had much luck getting that (and it's myriad dependencies) to
> build.

I also skip the gcc test suite, as the dependencies and infrastructure
are a bear.

>
> Are you talking about getting the normal gcc test suite to work, or
> are you talking about building a distribution of some other software
> that will test/verify the multilib bits?
>

The latter.  For example, take two trivial C programs:

- somelib.c, to be built 32-bit
- driver.c, with a dependency on somelib.o (main() is in driver.c)

so the two tests are:
1- can somelib.o be built as a 32-bit object on a 64-bit system --
this isn't a common use case, by the way; it's just needed for the
next test
2- can driver be built on a 64-bit platform and 'do the right thing'
with the 32-bit somelib.o -- this is the common use case I've seen our
users encounter (ie, they've gotten a 32-bit library from some vendor,
and now they need to use it on a 64-bit platform)

I've had user reports of some awkwardness linking things like that
correctly, and I'm experimenting with some different specfile
modifications & gcclib layouts to see if there is an easier way to
support compiles with vendor-supplied 32-bit libs.  Adding additional
link directives is the usual mechanism I recommend, but I'd like to
see if we can get the specfile & gcclib configured so that the search
path mechanism finds the right libraries automatically.  Suggestions
welcome there.

> In general, the efsdeploy build system avoids the need for using the
> multilib stuff by providing both 32 and 64 bit native tool chains,
> that produce the right default output.   In practice, that means you
> very rarely build a 32 bit binary using the 64 bit compiler, or vice
> versa (that does work, BTW, on some platforms).
>
>> 3- I notice fortran just got removed.  Do people actually need
>> Objective C?  (ie, --enable-languages objc)
>
> I have no idea -- I've always built gcc with those languages, and
> added fortran when it worked, an wasn't a PITA to build.   In the 25
> years I've been building gcc for people, I've very rarely had requests
> for fortran.  I can count them on one hand, I think.

It's rare in my experience as well, although I recently had to build
the ATLAS stack for someone (including LAPACK), and the user wanted
some fortran bits turned on, so it was necessary there.

Thanks,
Steven
_______________________________________________
EFS-dev mailing list
[email protected]
http://mailman.openefs.org/mailman/listinfo/efs-dev

Reply via email to