[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
Re: [equinox-dev] Class visibility without declared dependency
- From: Stephan Herrmann <stephan.herrmann@xxxxxxxxx>
- Date: Sat, 20 Jul 2013 22:33:01 +0200
- Delivered-to: email@example.com
- User-agent: Mozilla/5.0 (X11; Linux i686; rv:17.0) Gecko/20130623 Thunderbird/17.0.7
On 07/20/2013 09:43 PM, Richard S. Hall wrote:
On 7/20/13 10:46 , Stephan Herrmann wrote:
In the old framework I could use the ClassLoaderDelegateHook
to make aspect classes visible to those base classes into which
I'm weaving calls to the aspect. This is necessary, because
only the aspect knows about its base, not vice versa.
For Luna with the new WeavingHook all I can see is
This does, however, not work for packages that are not exported
from the aspect bundle. In fact most of the aspects are never
exported, because they're not meant to be called from source code.
If you manipulate some bundle's code and that code ends up with a dependency on some other external code, you might as well accept
the fact that it does become a true run-time dependency (i.e., a dynamic import) of the manipulated bundle. It doesn't make sense to
pretend that it isn't a dependency just because the dependency was introduced at run time. The framework has to handle it like any
I see your point, but it still is an incompatible change.
With our weaver on top of the old adaptor hooks an aspect bundle
does not have to declare this dependency.
I didn't even mention that we have an additional mechanism
called "forced-export", which allows an aspect to bind to
a non-exported class. That mechanism requires to add exports
at load-time. It's a pretty big hammer and fortunately rarely
needed, but I've seen situations where it rescued a project.
If you want to avoid that other bundles import the package in question without knowing that they shouldn't, then just introduce a
mandatory attribute on the exported package.
What is "mandatory"? Do you mean x-friends?
Additionally, in the old hook I had the option to expose individual
classes, not packages. I didn't yet use this power, but to enforce
that the synthetic dependency is indeed only ever used for the
synthetic aspect calls this would be valuable in the future.
So if a class-based mechanism is available I'd always prefer that
over a package based mechanism.
Any ideas how I could make this work in the new framework?
In OSGi, the atomic unit of dependencies is the package, there is no other way. The only wiggle room is with the include/exclude
directives on an export, but that isn't exactly the same.
I know that a manifest cannot import/export individual classes,
but a class loader hook can.
Currently, Object Teams can not run on the Luna framework and
for me this is a quite scary situation. Noticing my panic,
Tom had specifically invited me to ask on this list for help
how I can migrate the existing implementation of OT/Equinox
from the old adaptor hooks to the new framework.
From your answer I basically read we should have never done
what we did. That wasn't exactly my question :)
Any other ideas?