| Re: [cross-project-issues-dev] JFace Generics |
|
Guys, Contributions are a wonderful thing in principle, and should be strongly encouraged in practice, but using that as an argument to deflect technical criticism about the merit of generifying the JFace APIs is highly suspect. Of course everyone is more than welcome to deflect the technical criticism with technical arguments to contrary, as some have done. It's an open community and debate is more than healthy. It's already clear from the debate how poorly generics and arrays play together, and how unfortunately poorly that's understood. I.e., sorry but you cannot do "new E[size]", you just can't create a typed array without knowing the java.lang.Class. That's a major problem. Note that if you have a raw list, and you know/believe it to contain only elements of some specific type Foo, you can cast it to List<Foo> and use it. But even if you know an IStructuredContentProvider only works with instances of Foo and only returns Foo elements, casting it to IStructuredContentProvider<Foo> and working with it, you'll likely find that the Object[] result of getElements isn't a Foo[] and doesn't work. E.g., org.eclipse.compare.structuremergeviewer.DiffTreeViewer.DiffViewerContentProvider might be assumed to conform to ITreeContentProvider<IDiffElement, IDiffContainer> but its currently implementation doesn't conform until it's actually reimplemented "properly". I should emphasize too that I'm not so naive to argue there is zero merit to generification, but let's face the facts, these APIs, rife with Object[] instead of List<...>. As such, it's effectively stone age design when viewed in that light. An attempt to turn this into a modern generic container framework in face of the widespread use of array types which will never play well with generics seems to me at least somewhat misdirected. For example, another scenario where the changes will be problematic is in org.eclipse.jface.viewers.ITreePathContentProvider. That generification will have less than no value unless TreePath itself is generified: instance of checks on the path segments will remain if TreePath isn't a generic container, yet the return types of the methods are of the content provider are constrained, and those must be respected via actual changes in the code, not just by sprinkling the appropriate <> here and there. Addressing the generification of TreePath involves handling org.eclipse.jface.viewers.TreePath.TreePath(Object[]), and there are those darned arrays again. This can't be changed to E[] can it? Generic code creates these arrays to create the TreePaths, and such code can't create an array of E. So the end result will be no static type safety in creating the instance of TreePath if it remains Object[], and hence failures will happen later on access to the segments, in bridge methods without source. Is there improvement in this? It's totally frustrating to me that the platform hasn't yet address the last round of generification on Java collections, e.g., org.eclipse.jface.viewers.IStructuredSelection.iterator() is still raw and more annoyingly, so is org.eclipse.core.runtime.IAdaptable.getAdapter(Class). Yet now the goal is to introduce more such changes that will of course remain unaddressed in the entire huge ecosystem of downstream APIs. How long into the future will that disparity continue to exist? Will that make Eclipse look modern? Does that address the "why I hate Eclipse" mentality of the actual users of Eclipse. Does it introduce interesting powerful new functionality? Worse still for EMF specifically, is that we're maintaining compatibility with Eclipse 3.6, so how to address changes that require compiling against Eclipse 4.4's APIs? I suppose we could build against those new APIs and test against a 3.6 runtime, but does that really work in the face of the bridge methods generated when using these APIs? Perhaps it does, and given that I doubt there is a single place in the EMF code base that can actually make use of anything other than <Object> and <?>, maybe there will be no bridge methods, and we'll only end up with more noise in the code... Working in a side branch would seem far better. Then the full impact can be assessed before we're all fully impacted, and we can actually assess whether this all pans out, rather than the current approach that simply assumes it will all pan out reasonably well. It's an assumption that is questionable. Regards, Ed On 29/08/2013 10:20 PM, Doug Schaefer
wrote:
On 13-08-29 3:49 PM, "Aleksandar Kurtakov" <akurtako@xxxxxxxxxx> wrote:----- Original Message -----From: "John Arthorne" <John_Arthorne@xxxxxxxxxx> To: "Cross project issues" <cross-project-issues-dev@xxxxxxxxxxx> Sent: Thursday, August 29, 2013 7:37:17 PM Subject: Re: [cross-project-issues-dev] JFace Generics You raise good arguments as always Ed. I will attempt to summarize and respond to some of the comments raised in this thread. Q: Should JFace be generified at all. As with most library generifications there are pros and cons that we can probably debate all day. I still think on balance it will be a net benefit for consumers. Some will not have homogeneous element types and will be stuck with Object, and therefore be no better or worse off than before. TreeViewer is the worst case, although I think it would be difficult and awkward to generify some viewer types but not others. My experience with most generification is that 90% of usage is adapted relatively painlessly and eliminates a tonne of casting, resulting in cleaner code. There are always some rough edges, especially where arrays are found, and suppressing warnings is sometimes necessary. There is certainly work involved for clients to adopt it, although I don't find that alone a compelling argument to do nothing at all. A good place to continue this general discussion would be the master bug report: https://bugs.eclipse.org/bugs/show_bug.cgi?id=402445 Q: Should we release incomplete work to a milestone I'm torn on this one. Certainly in the past our project adopted a "we won't accept your contribution until it is perfect" approach. There is no doubt this benefits consumers, but I think it has strongly discouraged all but the most die-hard contributors to the project. Given the number of remaining active contributors, I strongly believe we need to find ways to make progress using the contributors we have, at the rate they are capable of contributing.Huge +1000 here. Eclipse platform is one of the projects that's hardest to get things into and I couldn't say that it's so much better than others like Gnome or KDE for example. The longer changes are delayed the more disruptive changes become. Even if certain things might be disruptive now this is the price to be paid for neglecting majority of contributors for years. Please keep opening the project - rejecting contributions is the road to oblivion. The question shouldn't be whether it benefits all API users but whether it breaks stuff. If it doesn't and if there is contributor willing to spend time on it he/she should be encouraged to contribute even if it will benefit single digit percentage of users. Growing contributors never starts with huge things, it works by piling stuff on and getting contributors used to workflow and guidelines in the project.Same here. You have to ask yourself whether rejecting contributions and evolution of the APIs has led to the reduced number of active contributors we have. I know it's more complicated than that, but please don't discount it.Alexander Kurtakov Red Hat Eclipse teamUndoubtedly this is a slower and bumpier road, but I think the risk is better than the alternative of stagnating because there are no contributors who can keep up with the old pace expected/demanded by the community. Certainly if there was risk of data loss or severe problems with tool usability, we would have avoided releasing in a milestone. In this case the worst case is compiler warnings which can be disabled if desired. I understand there are some who consider warnings to be catastrophically bad, but the community compiling against our milestones is a tiny fraction of the overall community. I believe all changes released were binary compatible and not changing behaviour. In the end I admit we got this one wrong. We misjudged the rate of progress on the work, and many committers being on holiday bogged down the review process. We should have kept the work in a branch and waited until at least after M1 to release. Some damage is already done, but we are now investigating moving the work into a branch. This will actually introduce a much worse problem for any client that had reacted to changes already, since it will result in compile errors. We'll send a separate warning note about this once we are sure on this path, so it doesn't get lost in this long-winded message. Q: I have some specific technical concerns about the approach taken so far. This is excellent feedback and I'm sure it will be incorporated into the work going forward. I wonder if we would have had such feedback if we hadn't released anything to master ;) Q: Your communication sucks. Yes it does. I think normally we wait until things are closer to completion before making big announcements. It is quite often that we have feature work in progress appearing in a milestone, that we hold off on announcing until a future milestone when it is more polished. In this case because the interim state already had significant source level impact this should have been communicated more widely. John From: Ed Merks <ed.merks@xxxxxxxxx> To: cross-project-issues-dev@xxxxxxxxxxx, Date: 08/29/2013 03:52 AM Subject: Re: [cross-project-issues-dev] JFace Generics Sent by: cross-project-issues-dev-bounces@xxxxxxxxxxx Hi, It's difficult to avoid an emotional outburst at this type of thing. I'm completely shocked that sweeping changes of this nature go unannounced on this mailing list. Sorry, a blog doesn't cut it... It's clear the current state is woefully incomplete. I.e., we have IStructuredContentProvider<E, I> but ITreeContentProvider is still raw. Of course it's clear that a tree rarely has uniformly typed children, so what is the plan for the completion of JFace's APIs? I question all this being committed to master in incremental stages like this... EMF is a sea of warnings with these changes, and eliminating those is weeks of work; work I can't begin because the changes are incomplete. And of course this affects all users of JFace, not just EMF, so the overall impact to the community is hard to calculate. The most disturbing part of all this is that I question whether it has even been well thought out. For example, the following change is highly disturbing: public interface IStructuredContentProvider<E,I> extends IContentProvider<I> { public E[] getElements(I inputElement); } Suppose I used it like this: public class GenericContentProvider<E, I extends List<E>> implements IStructuredContentProvider<E, I> { @Override public void dispose() { } @Override public void inputChanged(Viewer<I> viewer, I oldInput, I newInput) { } @Override public E[] getElements(I inputElement) { return (E[])inputElement.toArray(); } } I.e., I have a generic content provider implementation class that for a List<E> input returns the elements of that list. But there is a warning in the code for "E[])inputElement.toArray();" and it's not something one can blithely ignore. It will never be possible to create a generic subclass of IContentProvider that leaves E unsubstituted by a concrete implementation class, because it will never be possible to create an E[] array. If you question this assertion, stop and ask yourself why java.util.Collection.toArray() if of type Object[] and not of type E[]? It's because it would not be possible to implement generic Collection implementation classes. I can't emphasize enough how disturbing I find the approach being taken here. We're all familiar with using generics, but implementing generic classes properly remains complex and tricky and what's being done in JFace doesn't just impact the use of generics, it forces us all to revisit our implementation classes. For example, perhaps someone can explain how org.eclipse.jface.viewers.ArrayContentProvider will be updated? Probably just to "class ArrayContentProvider implements IStructuredContentProvider<Object, Object>" I would imagine, but that's not terribly useful is it? I imagine the overall impact on the community is to make sweeping pointless changes of precisely this nature. But suppose I even have a nice implementation class where I wanted IStructuredContentProvider<Foo, Bar>, my current implementation of getElements is probably wrong and would need to change to return Foo[]. But of what value is that? ContentProviders are generally just passed to a generic viewer, which uses it in a context where the types don't matter. So what's the benefit? Sorry to be so extreme in my opinion, but I would go further and argue that it's hard to imagine a significant set of scenarios where the new APIs are helpful even if this generic array issue wasn't just plain wrong or a horribly bad idea... I could go on and on, but as I said, it's hard to remain unemotional about this... Regards, Ed On 29/08/2013 7:46 AM, Lars Vogel wrote: Hi Eike, this is a GSoC done by Hendrik (cc) and was announced on PlanetEclipse. See http://blog.vogella.com/2013/06/17/google-summer-of-code-at-eclipse-2/ . John Arthone and I are the mentors for this project The work is still ongoing so far the ComboViewer and the TableViewer have been converted as well several basic classes. We currently don't know if we can generify IStructuredSelection. Input is very welcome, the umbrella bug is https://bugs.eclipse.org/bugs/show_bug.cgi?id=402445 and every code change is pushed to Gerrit. Look in https://git.eclipse.org/r/#/q/platform/eclipse.platform.ui,n,z for Hendrik (so don't know how to narrow the query down to Hendrik only). Best regards, Lars 2013/8/29 Eike Stepper < stepper@xxxxxxxxxx > Hi, After updating my target platform to Luna I noticed that my workspace is full of raw type warnings caused by changes in JFace to generify its APIs, for example Viewer. But the changes look incomplete, e.g., ViewerDropAdapter.getViewer() is still a raw type. Can we expect more such changes, e.g., IStructuredSelection? Has there been an announcement for these changes? Is there any advice on how to adjust our code? What about other Eclipse APIs, such as IAdaptable.getAdapter(Class), will it be changed to <T> T getAdapter(Class<T>)? Cheers /Eike ---- http://www.esc-net.de http://thegordian.blogspot.com http://twitter.com/eikestepper _______________________________________________ cross-project-issues-dev mailing list cross-project-issues-dev@xxxxxxxxxxx https://dev.eclipse.org/mailman/listinfo/cross-project-issues-dev _______________________________________________ cross-project-issues-dev mailing list cross-project-issues-dev@xxxxxxxxxxx https://dev.eclipse.org/mailman/listinfo/cross-project-issues-dev _______________________________________________ cross-project-issues-dev mailing list cross-project-issues-dev@xxxxxxxxxxx https://dev.eclipse.org/mailman/listinfo/cross-project-issues-dev _______________________________________________ cross-project-issues-dev mailing list cross-project-issues-dev@xxxxxxxxxxx https://dev.eclipse.org/mailman/listinfo/cross-project-issues-dev_______________________________________________ cross-project-issues-dev mailing list cross-project-issues-dev@xxxxxxxxxxx https://dev.eclipse.org/mailman/listinfo/cross-project-issues-dev_______________________________________________ cross-project-issues-dev mailing list cross-project-issues-dev@xxxxxxxxxxx https://dev.eclipse.org/mailman/listinfo/cross-project-issues-dev |