Dani,
As I've made pretty clear in the past, I'm a huge non-fan of
this effort. I find it ironic that the platform is rife with
raw types (List), and rather than investing an effort to
eliminate those, effort will be invested on generifying things
that just aren't screaming to be generified, and to do so in a
context that is heavily dominated by Object[], which interacts
exceedingly poorly with generics. Thought I've made that
argument already, I think it bears repeating...
Anywhere you see something like T[] where T is a type
parameter, you already have to get suspicious. Of course we
see this in java.util.Collection:
<T> T[] toArray(T[] a);
But note that T is derived from the argument, and argument my
not be null, so even in an erased runtime environment we can
determine a reasonable T from "a", so suspicion alleviated and
that's why we can have generic collection implementations...
Note that however nice it would have been that toArray()
method looked like this:
E[] toArray()
rather than
Object[] toArray
because you can't implement this generically, unless you
provide some object with the necessary
runtime information
about E to the constructor of an implementation class...
So consider this:
public interface IStructuredContentProvider extends
IContentProvider {
public Object[] getElements(Object inputElement);
}
there's no relationship between the input element's type and
the type of the array, so if someone proposes
public interface IStructuredContentProvider<X, Y>
extends IContentProvider<X> {
public Y[] getElements(X inputElement);
}
I'm going to be very suspicious. What this tells me is there
must be some sensible way of being sure that I'll be getting
back a real Y[] instance and not an Object[] instance. What
could that sensible way be?
Of course directly implementing that interface in a sensible
way is a good solution, but what about generic solutions?
Consider org.eclipse.jface.viewers.ArrayContentProvider (and
note that EMF generally has
only content provider
implementations analogous to this). This existing
implementation just don't work. You can just forget about
org.eclipse.jface.viewers.ArrayContentProvider.instance, you
can just deprecate
org.eclipse.jface.viewers.ArrayContentProvider.getInstance(),
and you'd better add a public constructor and deprecate it
while your at it, because this implementation
public Object[] getElements(Object inputElement) {
if (inputElement instanceof Object[]) {
return (Object[]) inputElement;
}
if (inputElement instanceof Collection) {
return ((Collection) inputElement).toArray();
}
return new Object[0];
}
simply doesn't work for collections. You'll need new
constructors and new getInstance methods each of which specify
the array type, either as java.lang.Class or as an array
prototype, as in the first form to toArray above. You'd have
to provide that even for the constructor that takes a
collection, just the collection instance will not suffice.
Great, so much for adding generics to JFace being erasure
compatible, counter to what was the case when Java's
collection library was generified. Nothing in the
collections library was deprecated and no new methods were
added. In other words, for JFace's changes, you can't just
turn off warnings about raw types, you must deal with the
deprecations and API changes. So if you're trying to maintain
something that works with old versions of Eclipse (i.e., EMF
is compatible with Eclipse 3.5), you're completely hosed. You
can't add generics, because you can't compile against an older
target platform that does have it, so you simply have to live
with a sea of raw type warnings (or turn them all off and lose
the value of even having such warnings). Also, you can't
start using the new methods instead of the deprecated ones, so
you have to live with that sea of warning as well, or just
turn them off too. Nor you can you exploit any of it to
provide value to clients (given the premise there is value),
because you can't reasonably address this ArrayContentProvider
problem without inflicting the same pain on the clients (who
actually have better things to do, go figure).
Even the premise that this effort has value is questionable.
Granted, someone writing their first content provider might
find it useful, if (and only if) it's one that's concrete and
if (and only if) it doesn't need to deal with several input
types that have no common super type (and even in that case
they'll still typically end up with instanceof tests to return
subtype-appropriate results). That's on the argument side of
the getElements. On the return type side, it's of no benefit
to the author; they just pass this content provider into a
generic viewer that doesn't care whether it's Object[] or X[].
So in fact it's just a burden with which one must conform.
So is the value of this whole exercise eliminating instance of
checks for the arguments of the provider implementations?
Does this tangible (but small) benefit justify the impact on
the long established community? Do we expect that community
to eliminate their deprecations and generify all their code?
(Sorry guys and girls, all your existing toArray() calls are
invalid, or, don't worry about it, just sprinkle <Object,
Object> everywhere.) I wonder, will JDT and PDE do that?
If not, why expect the rest of the community to do it? And if
you don't expect that, what exactly are you expecting will
come from this?
I suggest folks carefully weight the benefits against the
disruptive nature of this type of change.
Regards,
Ed
On 30/07/2014 11:42 AM, Daniel Megert wrote:
Just for the records, here are
some constraints that I required in order to agree to
continue that work:
- Some stuff just doesn't
make sense to be generified because it often contains
various kinds of objects, e.g. (tree) viewers. See also http://dev.eclipse.org/mhonarc/lists/platform-ui-dev/msg05459.html.
- If generified types
cannot be plugged together unless everything is again just
Object or Class, it's not worth to generify those types.
- The generified code must
be in a shape so that clients can start to fix their code
by invoking Refactor > Infer Generic Type Arguments...
This needs to be validate on existing Platform UI code.
Dani
From:
Lars Vogel <lars.vogel@xxxxxxxxx>
To: cross-project-issues-dev@xxxxxxxxxxx,
Jeanderson Candido <jeandersonbc@xxxxxxxxx>,
Hendrik Still <Gamma32@xxxxxxxxx>
Date:
30.07.2014 11:23
Subject:
[cross-project-issues-dev]
Information
about the "Generifing JFace viewers" project
Sent by:
cross-project-issues-dev-bounces@xxxxxxxxxxx
Hi,
as some of you probably remember, the
platform.ui team started a GSoC project last year to
generify the JFace viewer framework. We (platform.ui team
together with John Arthone and Dani Megert) decided that
it is worth to finish this project and started a new GSoC
project.
Jeanderson Barros Candido (cc) is working on
this project with Hendrik Still (cc) (GSoC student from
last year) and me as mentor.
I personally think the work looks already
very good and plan to integrated it soon into the master.
We are trying to learn from the experience from last year,
therefore:
- We plan to integrate it as a whole, not
piece wise so people can fix warning messages created by
this change
- We reworking the JFace
snippets and tests at the same time to have a first
proof-point
- We plan to use it for platform
views to validate that it works
Of course generifying an existing API, will result in
certain limitations and some suggested a complete rewrite
of the JFace viewer framework but this is currently not
the scope of this project.
The implementation is currently done at
Github: https://github.com/jeandersonbc/eclipse.platform.ui and we do our planning in https://github.com/jeandersonbc/gsoc14-eclipse-planning.
If someone wants to test the new
implementation and provide feedback, please let us know.
Best regards, Lars_______________________________________________
cross-project-issues-dev mailing list
cross-project-issues-dev@xxxxxxxxxxx
To change your delivery options, retrieve your password,
or unsubscribe from this list, visit
https://dev.eclipse.org/mailman/listinfo/cross-project-issues-dev
_______________________________________________
cross-project-issues-dev mailing list
cross-project-issues-dev@xxxxxxxxxxx
To change your delivery options, retrieve your password, or unsubscribe from this list, visit
https://dev.eclipse.org/mailman/listinfo/cross-project-issues-dev