On 05 Nov 2013, at 12:41, Ward Poelmans wrote:
> On Tue, Nov 5, 2013 at 11:16 AM, Kenneth Hoste <[email protected]> wrote:
>>> On Mon, Nov 4, 2013 at 6:35 PM, Fotis Georgatos <[email protected]> wrote:
>>>> * toolchain-less builds; important for debuggers, performance tools and
>>>> all non-ABI exposing software
>>>
>>> +1
>>
>> Are you asking for easyconfigs that use a dummy toolchain? I'm OK with
>> merging them in, if PRs are issued for them.
>>
>> This fits in the minimal-vs-full toolchain discussion, see also
>> https://github.com/hpcugent/easybuild-wiki/pull/11 .
>>
>> I'm not sure if there's anything to discuss here. It's not like we oppose
>> dummy-toolchain easyconfigs, we just don't think it's a good idea (which
>> doesn't mean we're not willing of merging them in).
>>
>> To also make them used by existing easyconfigs, we probably need
>> https://github.com/hpcugent/easybuild-framework/issues/741 to be tackled
>> first.
>> That would give everyone the freedom to go for one of both schemes (with
>> dummy toolchain builds to be an extreme case of minimal toolchains).
>
> Well, what is the purpose of the dummy toolchain? Let's take gdb as an
> example: it needs a c/c++ compiler but which one or which version does
> not really mater. Is the dummy toolchains the best approach here?
How sure are you that the GDB binary you obtain is really independent from the
compiler that was used?
Any dynamically linked libraries provided by GCC involved? If so, what happens
if you load a GCC module that is older than the system GCC which was used for
building GDB (if a dummy toolchain is used)?
If the GDB binary is truly 'stand-alone', then yes, using a dummy toolchain for
it makes sense.
Note that this will require you to explicitly add make options like "CC=gcc" in
the GDB easyconfig, since we hard set those to nonsense ('dummyCC' or somesuch)
with a dummy toolchain currently (to avoid that people use it by accident).
> I would be interested to hear what the current status is of the
> toolchains soup: you want to build something with a certain compiler
> but first you need to update several deps to the correct toolchain
> version, create new easybuilds because they only exists for gcc and
> not ictce (or vice versa), ... Is this something the v2.0 format will
> solve?
Good question, let's discuss this during the conf call. Added it to the agenda.
K.
