I have an article available for publication for national audiences only. This means that the article should either be translated into a local language (anything but English), either distributed to a closed community (subscribers only).
If anyone is interested, please contact me. And if you want your favorite magazine to publish something about PostSharp, forward them this page :-D.
P.S. The article is great!
In my previous post I claimed that it was not possible, in general, to merge assemblies enhanced by PostSharp, because it would break the assembly references stored as strings in aspect serialization data.
I did not realize that the BinaryFormatter provides a way to solve our problem. It is indeed possible to provide its own SerializationBinder, whose role is precisely to return the System.Type corresponding to an assembly name and a type name. We can provide a custom binder to redirect merged assemblies.
This is the feature I have added to build 184.108.40.2061 available for download in the builds section; I could still do it in 1.0 since it is a no-risk feature, breaking nothing existing (if you don't use it, nothing can go wrong).
I have added a new class PostSharp.Laos.LaosSerializationBinder. You can set the static property Current to your own implementation of SerializationBinder. If you don't set it, nothing is changed. The LaosSerializationBinder class it itself derived from SerializationBinder and implements the feature we need: assembly retargeting. So all you have to do is to create an instance of this class, set up retargeting policies, and assign it to LaosSerializationBinder.Current.
The following code sets up PostSharp Laos so that references to 'MergedPostSharpLib.dll' are retargeted to 'MergedPostSharp.exe' (supposing that the library has been merged into the executable):
LaosSerializationBinder binder = new LaosSerializationBinder();
LaosSerializationBinder.Current = binder;
The only thing you have to really take care about is to set up the binder before the first aspect is deserialized. Aspect deserialization is triggered from static constructors of enhanced classes, so you may need to initialize the weaver very soon, and eventually to refactor your "Main(string args)" method.
Apart from that, it should work very well.
Combining PostSharp with ILMerge seems a recurring demand. Indeed, many developers don't welcome the multiplication of dependent assemblies as PostSharp.Public.dll and PostSharp.Laos.dll, and ILMerge seems the dreamed solution to that. Unfortunately, it will work only in rare cases.
The Simple Case
Say you have an assembly named MyProgram.exe with aspects defined in this assembly. The application is formed of the following assemblies:
In this particular case, it is valid to merge all three assemblies into a new assembly equally named MyProgram.exe.
However, if you modify the name of the merged assembly, it will cease to work.
Remember that the aspects are serialized into the assembly. It uses the BinaryFormatter to do the job. Unfortunately for our purpose, this formatter stores the full type name of aspects, including the assembly name. So when you rename the assembly, you break the link between the aspect serialization and the underlying types, and you get an exception during deserialization.
The same occurs when you have aspects in one of the dependencies you merge. Say your program references MyLibrary.dll, whose some classes have been enhanced by PostSharp Laos. If you merge it into MyProgram.exe, the serialization of aspects previously defined in MyLibrary.dll will be broken.
The problem clearly lays in the use of the BinaryFormatter to do the job. We could eventually use another serializer that does not store the full type name.
This feature (pluggable serializers) was actually present in former builds of what is now PostSharp 1.5, but I have removed the feature because it became useless. If there is some interest, I could restore it. It very much depends on the feedback I will get.
How is this a critical feature for you?
Now that PostSharp 1.0 is pretty stable and anyway virtually frozen, I am pleased to announce that the first community technical preview of PostSharp 1.5, the current living branch of PostSharp, is now available for download.
This CTP makes a great leap forward in supporting more platforms:
- Novell Mono. PostSharp now run on Mono and has been tested both in the Windows and the Linux implementation. PostSharp is now truly cross-platform, since the same PostSharp executable can now run on any platform.
- Microsoft Silverlight 2. Who said aspects are only for server applications? Thanks to PostSharp 1.5, you can now apply aspects to your projects targeting Silverlight 2.
- Microsoft .NET Compact Framework 2.0.
Enlarging PostSharp support to these platforms required the possibility to read and transform assemblies without loading them in the CLR. Therefore, one of the hidden new features of PostSharp 1.5 is the ability to read assemblies directly from disk.
Another hidden feature is the support for assembly retargeting, which happens when an assembly reference is resolved to an assembly of higher version than was required. This feature, supported by the .NET framework, was missing in PostSharp. This first CTP offers a first yet incomplete solution to this issue.
The documentation and samples have been updated to include support for Silverlight and the Compact Framework.
Note that PostSharp 1.5 is not fully compatible with PostSharp 1.0. End-users should not see the difference, but if you played with PostSharp Core or with some advanced features of PostSharp Laos, your code will need some changes.
Most people think of PostSharp as a new product, but PostSharp is actually near to 4 years old. The first bits were indeed developed in September 2004. It took roughly 2 years to develop the product and 2 years to stabilize days. Attentive users have maybe noticed the last RC of PostSharp has the full version number 220.127.116.113. 10 means that it is the 10th release of the 1.0 branch. 403 is the revision number of the root of that branch in source control.
Today the 1.0 branch is very stable. There has been no major change and so significant feature addition for one year. It is being downloaded more than one thousand times per month (still quite a modest number). So the future of the 1.0 branch is clear: fix bugs and declare the branch stable. More a marketing operation than a technical one.
The 1.0 branch will remain a legacy one. Bugs will continue to be fixed, support will be provided. So it is definitively a product you can rely on.
In October 2007, I forked the 1.0 branch into 1.1. Unfortunately, I had to interrupt the work, principally because much work was still required on 1.0, and on the web site. Good news is that I resumed work around last April and made good progresses.
Since this new branch contains major new features, I chose to number it 1.5.
The objectives of this branch are triple:
- Enlarge supported platforms to Mono, Compact Framework and Silverlight 2 [CTP 1, July 2008].
- Implement frequently asked feature [CPT 2, August-Sep 2008]
- Allow aspects to be inherited (for instance, if you apply an aspect on an interface, the aspect will be applied to all types implementing it) [CPT 3, Sep-Oct 2008].
After the cycles of CTP, there will be a cycle of Beta releases during which features will be added according to the feedback got during CTP. Then release candidates and hopefully RTM.
A characteristic of 1.5 is that it won't be fully compatible with 1.0. End users (those writing aspects using PostSharp Laos) should not see a big difference and most of their existing code should still work, but some minor changes are to be expected for Core developers. Not a big deal, but it's better to know it.
Design by Contract
PostSharp 1.5 will provide the basic bricks necessary to build a real design-by-contract (tm, I know) framework. I still don't know exactly what it will contain, but you can obviously expect the non-nullable references and similar value constraints, including a mean to do your own validators without the burden of PostSharp Laos. But the framework should be more ambitious than just value checking. Invariant checks and defense (eventually in a multithreaded context) are on the wish list.
A first version will be available in November 2008.
I have no plan for 2009, but I guess it will be determined by your feedback on these new features.
I am delighted to announce that I will present PostSharp (officially, aspect-oriented programming and design by contract) in Baden-Baden, Germany, on November 11, 2008 at the prio.conference 2008 Home Page, dedicated to software architecture.
Germany is already the country where PostSharp is arguably the most popular, certainly because of the influence of the early enthusiast Ralf Westphal.
Why design by contract, you may ask? Well, you'll discover it soon on this blog...
The last release candidate was unfortunately not the final one. Last Friday I silently released a new candidate, and since there seems to be no obvious error, it's not time to announce it (I had to retract a previous tentative to public RC4 for this reason).
This new RC corrects fourteen issues, two of them required non-trivial changes requiring a new RC.
Maybe the best news for all is that it is now possible to apply multiple OnMethodInvocation aspects on a single method, at least with target-site weaving. I have been shamefully so far too lazy to reconsider a limitation that no longer any rationale.
I was however quite satisfy to notice that most bugs had pretty small coverage, so I consider already the product as very stable.
Unless you guys come with a new high-coverage bug requiring non-trivial refactoring, this will be the final release candidate.
Dependency Injection is Dead! is the somewhat provocative title of an interesting post of Simon Ince.
Simon indeed proposes another approach: AOP-powered Dependency Resolution. The idea is to use aspects on fields or properties to resolve the dependencies on-demand, when the property or field value is retrieved. Technically, it would be enough to turn the old good [Dependency] custom attribute into an aspect.
I won't copy-paste Simon's stuff, so please read from the horse's feather on his blog.
A refreshing view on a topic that tends to become sanctified and therefore unchallenged in our community.
Even if this approach does not solve all the use cases of Dependency Injection, covers a lot of them in quite an elegant way. While it would be theoretically easy to use this approach with dependency injection frameworks (because dependency resolution is obviously one the major components of a DI framework, the other big one being the object builder), I noticed that Unity does not offer the functionality at the required level of granularity. A pity, which I have notified in illo tempore to the Unity team. Maybe some day wind will turn...