[jsr294-modularity-eg] Problems for JSR 294 to address

[Alex Buckley]

.NET is not Java.

Compiled types in .NET live in non-executable modules (.netmodule 
files). To be executed, modules must be packaged into assemblies 
(.dll/.exe files).

This packaging makes "internal" a bit tricky. "internal" is specified as 
"accessible from the same program", where a program is merely a bunch of 
source files. A C# compiler must take it on trust that the .netmodule 
which accesses an "internal" artifact will be packaged in the same 
assembly as the .netmodule which declares the artifact.

This is not unreasonable, since .netmodules MUST be packaged. But if 
packaged "wrongly", code accessing an "internal" artifact which the 
compiler promised was accessible will fail at runtime. You do not want 
language features to behave like this. There is effectively a hole in 
the C# spec caused by the greater CLI spec stratifying loadable code 
into non-executable and executable.

Java is different. A JVM does not require classfiles to be packaged in 
any way before executing them. Nor does the reference implementation of 
a Java compiler require classfiles to be packaged before compiling 
against them. These things are not going to change. Since module-private 
accessibility is a language and VM construct, it cannot depend on the 
future packaging of code still being compiled or compiled against.

(What it can depend on is an API consulted by a Java compiler to 
discover whether module M version X and module M version Y have "equal" 
versions. The API implementation would be honor-bound to give consistent 
answers, similar to classloaders. We'll get to that API soon, I hope.)

This boils down to saying the module-private accessibility and module 
membership go hand-in-hand for me. In any case, the first thing a 
programmer will ask when seeing a module-private member is: "What module 
does this member belong to?". The answer should NOT be "it depends on a 
compiler flag or invocation of jar".

I will shortly send a Java language grammar and classfile format which 
reflect these views.

Now, I know this issue is not going to go away. I therefore propose that 
IN ADDITION to encoding module membership in source, it should be 
possible for the host system to determine module membership. If you 
REALLY want the answer above to be "it depends on a compiler flag or 
invocation of jar", you can do that. But a Java compiler/VM must defer 
to the membership in a source/class file if it exists.

Is that enough of a weakening of my requirements #1 and #2?

Alex

Peter Kriens wrote:
> Hmm. .NET shows that you do not have to know all the module boundaries 
> during compilation. 'internal' indicates inside the assembly and the 
> source assembly is clearly not known during compilation. This concept 
> looks attractively simple and is in line with common practice.
> 
> For 294, during compilation, the JAR files and project source can easily 
> be used inside the compiler for the scope. During runtime, the module 
> system can set this boundary.
> 
> However, I also like a module concept that is defined in the language. 
> This will be more complex (and probably should be hierarchical) but it 
> should be not attempt to define the deployment artifact nor 
> dependencies. That is, it should be more like namespaces.
> 
> Kind regards,
> 
>     Peter Kriens
>     
> On 28 jan 2009, at 03:56, Alex Buckley wrote:
> 
>> I am happy to discuss module-private accessibility first, in that it 
>> addresses the limitations of packages on which we all agree.
>>
>> The issue we cannot ignore is that to check module-private 
>> accessibility at compile-time means knowing the modules to which the 
>> requesting and requested types belong. If I invoke:
>>
>> javac Foo.java Bar.java
>>
>> where Foo.java refers to Bar, then a compiler must be able to compute 
>> at least the module name to which Foo belongs as Foo is compiled, if 
>> Bar is found to be module-private as it is compiled. There are no 
>> deployment artifacts around yet. It shouldn't be a surprise that 
>> Foo.class and Bar.class are subsequently emitted to encode their 
>> module membership.
>>
>> So with module accessibility comes module membership. Module 
>> dependencies are a quite different topic which we'll get to soon 
>> enough, and they are the place where minimizing new concepts is 
>> essential.
>>
>> Alex
>>
>> Peter Kriens wrote:
>>> I think it would make sense to work from first principles
>>>   1. What problem are we trying to solve?
>>>   2. How do we scope this problem to an acceptable area for all?
>>>   3. What requirements should the solution fulfill
>>> I would like to have a short discussion about the problem so we are 
>>> sure we are all on the same page. I can start off with the problem I 
>>> think we are trying to solve.
>>> Modularity
>>> Modularity is the art of encapsulation and hiding. A module limits 
>>> the amount of information outside its boundaries. This reduces the 
>>> overall complexity of a system because it becomes possible to reason 
>>> (and change) locally about a module instead of understanding the 
>>> whole system, and it provides a local namespace that is easier to 
>>> work with for humans than a global namespace.
>>> Java provides modularity in many places. A type encapsulates its 
>>> fields and methods and selectively exposes them to other types with 
>>> the access modifies: private, protected and public. A package 
>>> encapsulates a number of types and resources. And last, but not 
>>> least, the class loader provides a class space that can be distinct 
>>> from other class spaces.
>>> The modularity that is enabled with class loaders has been exploited 
>>> by many. However, the Java Language has little to say about class 
>>> loaders; it is outside the scope of the language.
>>> So what problem do we need to solve, seeing that Java provides 
>>> already so many forms of modularity? There is a sense in the 
>>> community that packages are too small for modules. Though the package 
>>> names hint at a hierarchy (and this is how they are found in a file 
>>> system), this hierarchy does not provide a preferential treatment for 
>>> children. Many programs consist of hundreds of packages. However, 
>>> except for the limited package concept, the programmer cannot 
>>> indicate that the visibility of an artifact should be limited to the 
>>> program, or part of the program.
>>> A large number of delivery formats have been created over the years, 
>>> most of the based on the JAR format (which is also not a part of the 
>>> language): WAR, EAR, OSGi bundles, midlets, xlets, and likely many 
>>> more proprietary formats. Usually, the modularity of these modules is 
>>> based on a rather simplistic hierarchical class loader model. That 
>>> is, all classes in ancestors are visible, but not any siblings or 
>>> their children. The language is moot on the point of this type of 
>>> encapsulation.
>>> Delivery modules contain types that depend on types in other delivery 
>>> modules. There is no uniformly agreed Java standard to model these 
>>> dependencies. During build time, the compiler is provided with a 
>>> linear list of JARs and the compiler picks the first type that 
>>> matches a name. Only the name is encoded in the class file, not the 
>>> originating delivery module. During runtime, the same process is used 
>>> to find classes, albeit in a hierarchical class loader model. This 
>>> has all so far been outside the Java Language.
>>> It is crucial that we distinguish the language/logic modularity and 
>>> the modularity based on the deployment artifact. Interestingly, in my 
>>> understanding, .NET makes such a distinction between the modularity 
>>> of a delivery unit (called an assembly) and the finer grained 
>>> modularity inside a delivery unit, represented by namespaces. The 
>>> "internal" keyword indicates that the artifact is visible only inside 
>>> an assembly and the "namespace" keyword provides a hierarchical 
>>> namespace.
>>> Therefore, one can identify two problems in the Java platform 
>>> concerning modularity
>>>   1. Packages are too limited for proper language/logic modularity
>>>   2. Lack of a runtime module concept that is related to deployment
>>>      artifacts (physical modularity)
>>> Looking at the proposed time frame (EDR by mid-April), attacking both 
>>> problems simultaneously seems rather unrealistic based on my limited 
>>> experience. I therefore suggest to start with problem #1: Packages 
>>> are too limited for proper language/logic modularity. What do you think?
>>> Kind regards,
>>> Peter Kriens
>>> On 27 jan 2009, at 05:31, Alex Buckley wrote:
>>>> Comments welcome.
>>>>
>>>> 1) Packages are typically arranged in hierarchies, but types can 
>>>> only be
>>>> used across different branches of the hierarchy by being made public,
>>>> which exposes implementation details too widely. Information hiding is
>>>> further reduced by interface members always being public.
>>>>
>>>> Non-solution: hierarchical package membership. Redefining existing
>>>> well-known semantics is always a bad idea, and this one is especially
>>>> complicated. There would need to be a way to stop package-private
>>>> artifacts from being accessible to subpackages, or else we would be
>>>> strengthening information hiding in one place only to weaken it
>>>> elsewhere. Also, there would need to be a way to configure the depth of
>>>> exposure of package-private artifacts, since artifacts in package A
>>>> should sometimes be accessible to A.B and not to A.B.C, and so on.
>>>>
>>>> 2) Dependencies of one type on another are expressed in source and
>>>> classfiles, but there is no standard way for a Java compiler or JVM to
>>>> interact with its environment to read and resolve dependencies. This
>>>> causes compile-time and runtime environments to differ, and complicates
>>>> any effort to version types available in the (compile-time or runtime)
>>>> environment.
>>>>
>>>> Non-solution: standardize the CLASSPATH. One list for all programs and
>>>> libraries in the JVM makes versioning almost meaningless, and packages
>>>> whose types occur in multiple CLASSPATH entries can behave poorly. (See
>>>> Peter Kriens' presentation at Devoxx 2008.)
>>>>
>>>> 3) Many packages have grown large over the years. They are effectively
>>>> impossible to refactor now, since it is binary-incompatible to rename
>>>> types and dangerous to "split" packages. The next-best option is to let
>>>> subsets be taken of existing packages, and deliver subsets
>>>> independently, but there is no mechanism in the Java language or JVM 
>>>> for
>>>> that.
>>>>
>>>> Non-solution: a mechanism for renaming packages and types outside the
>>>> Java language. This would require updating Java's model of binary
>>>> compatibility, and would make source code incomprehensible.
>>>>
>>>> Alex
>>>> _______________________________________________
>>>> jsr294-modularity-eg mailing list
>>>> jsr294-modularity-eg at cs.oswego.edu 
>>>> <mailto:jsr294-modularity-eg at cs.oswego.edu>
>>>> http://cs.oswego.edu/mailman/listinfo/jsr294-modularity-eg
>>> ------------------------------------------------------------------------
>>> _______________________________________________
>>> jsr294-modularity-eg mailing list
>>> jsr294-modularity-eg at cs.oswego.edu
>>> http://cs.oswego.edu/mailman/listinfo/jsr294-modularity-eg
>> _______________________________________________
>> jsr294-modularity-eg mailing list
>> jsr294-modularity-eg at cs.oswego.edu
>> http://cs.oswego.edu/mailman/listinfo/jsr294-modularity-eg
> 
> _______________________________________________
> jsr294-modularity-eg mailing list
> jsr294-modularity-eg at cs.oswego.edu
> http://cs.oswego.edu/mailman/listinfo/jsr294-modularity-eg