We check our software architecture by some ArchUnit rules.
One of this is the test for our layered architecture.
This works fine for methods.
If we access a method of layer3 from layer1 we will get an exception.
But if the access a field declared in layer3 from layer1 this won't throw an exception.
.layer("layer1").definedBy("com.acme.layer1")
.layer("layer2").definedBy("com.acme.layer2")
.layer("layer3").definedBy("com.acme.layer3")
.whereLayer("layer3").mayNotBeAccessedByAnyLayer()
.whereLayer("layer2").mayOnlyAccessedByLayers("layer3")
.as("Respect the layered architecture");
This won't throw an exception if we import a field from layer3 to a layer1 class:
package com.acme.layer1
import static com.acme.layer3.SOME_LABEL
public class x {
...
}
Our expectation is that also accessing fields from layer3 in any other layer should throw an exception.
Or is there another way to check?
My answer depends on the assumption that com.acme.layer3.SOME_LABEL is a constant expression like public static final String SOME_LABEL = "..."
Compile-time constants are inlined during compilation. If your code looks like
String label = SharedConstants.SOME_LABEL;
then the compiled bytecode contains the exact value of SharedConstants.SOME_LABEL. There remains no reference to this field in the bytecode. (The Static Final Inline Trap)
ArchUnit gathers all information by analyzing the bytecode, see also the ArchUnit User Guide. Because there is no information about SharedConstants.SOME_LABEL in the bytecode, ArchUnit doesn't know about this access.
In summary: This is a limitation of ArchUnit and all other libraries which only depend on the bytecode.
Related
To externalize UI strings we use the "Messages-class" approach as supported e.g. in Eclipse and other IDEs. This approach requires that in each package where one needs some UI strings there has to be a class "Messages" that offers a static method String getString(key) via which one obtains the actual String to display to the user. The Strings are internally accessed/fetched using Java's Resources mechanism for i18n.
Esp. after some refactoring - we again and again have accidental imports from a class Messages from a different package.
Thus I would like to create an archunit rule checking whether we only access classes called "Messages" from the very same package. I.e. each import of a class x.y.z.Messages is an error if the package x.y.z is not the same package as the current class (i.e. the class that contains the import)
I got as far as this:
#ArchTest
void preventReferencesToMessagesOutsideCurrentPackage(JavaClasses classes) {
ArchRule rule;
rule = ArchRuleDefinition.noClasses()
.should().accessClassesThat().haveNameMatching("Messages")
.???
;
rule.check(classes);
}
but now I got stuck at the ???.
How can one phrase a condition "and the referenced/imported class "Messages" is not in the same package as this class"?
I somehow got lost with all these archunit methods of which none seems to fit here nor lend itself to compose said condition. Probably I just can't see the forest for the many trees.
Any suggestion or guidance anyone?
You need to operate on instances of JavaAccess to validate the dependencies. JavaAccess provides information about the caller and the target such that you can validate the access dynamically depending on the package name of both classes.
DescribedPredicate<JavaAccess<?>> isForeignMessageClassPredicate =
new DescribedPredicate<JavaAccess<?>>("target is a foreign message class") {
#Override
public boolean apply(JavaAccess<?> access) {
JavaClass targetClass = access.getTarget().getOwner();
if ("Message".equals(targetClass.getSimpleName())) {
JavaClass callerClass = access.getOwner().getOwner();
return !targetClass.getPackageName().equals(callerClass.getPackageName());
}
return false;
}
};
ArchRule rule =
noClasses().should().accessTargetWhere(isForeignMessageClassPredicate);
To my surprise I have learned today that the following works just fine in Groovy:
import java.util.concurrent.*
def atomicBool = new atomic.AtomicBoolean(true)
i.e. after the star import, I can use a 'partial' package to refer to java.util.concurrent.atomic.AtomicBoolean.
Obviously, the same is not valid in Java:
import java.util.concurrent.*;
public class Sample {
public static void main(String[] args) {
// compiler error: 'atomic' cannot be resolved to a type
new atomic.AtomicBoolean(true);
}
}
So it seems that Groovy's idea of a package is similar to C++ (or C#) namespaces in this regard.
Question to the Groovy experts:
Is this by design or is it a (potentially unintended) side effect of the way the interpreter treats star imports?
If it is by design, can you point me to a section in the documentation or language specification where this behavior is documented? (There is no mention of this in the documentation on Star Import and neither in the language spec as far as I can tell or at least I couldn't find anything.)
Based on Groovy source code, this behavior seems to be intended. Before we dig deeper into Groovy internals you have to be aware of one thing - Groovy compiles to a bytecode that can be represented by a valid Java code. It means that Groovy code like the one form your example actually compiles to something like this (without compile static and type checked transformations):
import groovy.lang.Binding;
import groovy.lang.Script;
import java.util.concurrent.atomic.AtomicBoolean;
import org.codehaus.groovy.runtime.InvokerHelper;
import org.codehaus.groovy.runtime.ScriptBytecodeAdapter;
import org.codehaus.groovy.runtime.callsite.CallSite;
public class test extends Script {
public test() {
CallSite[] var1 = $getCallSiteArray();
}
public test(Binding context) {
CallSite[] var2 = $getCallSiteArray();
super(context);
}
public static void main(String... args) {
CallSite[] var1 = $getCallSiteArray();
var1[0].call(InvokerHelper.class, test.class, args);
}
public Object run() {
CallSite[] var1 = $getCallSiteArray();
AtomicBoolean atomicBool = (AtomicBoolean)ScriptBytecodeAdapter.castToType(var1[1].callConstructor(AtomicBoolean.class, true), AtomicBoolean.class);
return var1[2].callCurrent(this, atomicBool);
}
}
As you can see this Java class uses full java.util.concurrent.atomic.AtomicBoolean import and this is actually what Groovy transforms your input source code to.
How does it happen?
As you may know Groovy builds Abstract Syntax Tree (AST) from the input source file and it iterates over all nodes (like expressions, variable definitions, method calls etc.) and applies transformations. Groovy uses class called ResolverVisitor that is designed to resolve types. When Groovy compiles your code it finds ConstructorCallExpression:
new atomic.AtomicBoolean(true)
It sees that the expected type of the object you are trying to create is atomic.AtomicBoolean, so ResolverVisitor starts resolving the type by calling resolveOrFail(type, cce); at line 1131.
It tries several resolving strategies that fail until it reaches method resolveFromModule at line 695. What happens here is it iterates over all star imports (single java.util.concurrent.* in your case), then it concatenates star import with the type name and it checks if qualified name created from this concatenation is a valid type class. Luckily it is in your case:
When type gets resolved, Groovy replaces initial type with this resolved valid type name in the abstract syntax tree. After this operation your input code looks more like this:
import java.util.concurrent.*
java.util.concurrent.atomic.AtomicBoolean atomicBool = new java.util.concurrent.atomic.AtomicBoolean(true)
This is what compiler gets eventually. Of course fully qualified name gets replaced with the import (this is what Java compiler does with qualified names).
Was this "feature" introduced by designed?
I can't tell you that. However we can read from the source code, that this happens on purpose and type resolving like this is implemented with intention.
Why it is not documented?
I guess no one actually recommend using imports that way. Groovy is very powerful and you are able to do many things in a many different ways, but it doesn't mean you should do so. Star imports are pretty controversial, because using star imports over explicit imports makes your code more error-prone due to possible class import conflicts. But if you want to know exact answer to this kind of questions you would have to ask Groovy language designers and core developers - they may give you straight answer without any doubts.
I have two constructors: normal ctor which initialises the object properly and a second default ctor for Hibernate which generates initialize fields warnings. What's the preferred way to get rid of the warnings?
package test;
public class Example {
String x;
public Example(String x) {
this.x = x;
}
Example() {
// Ctor for Hibernate, warnings generated here.
}
}
You didn't mention looking in the documentation, so I'm not sure whether you have done so. The Checker Framework manual contains a chapter titled "Suppressing warnings", which might contain all the information you need.
The most common approach is to write a #SuppressingWarnings annotation, which is the standard way to suppress warnings from the Java compiler.
You should write it on the smallest program element possible (such as a local variable declaration rather than the whole constructor or class), and you should supply the most specific key possible. The reason is to avoid accidentally suppressing more warnings than you intended.
I need to build a grammer containing a cross reference, which may be invalid, i.e. points to a nonexisting target. A file containing such a reference should not yield an error, but only a warning. The generator would handle this as as a special case.
How can I do this with XText?
It's not possible to create valid cross references to non-existing targets in EMF.
I would suggest to go with EAttributes instead of EReferences:
Change the feature=[EClass|ID] by feature=ID in {YourDSL} grammar.
Provide a scope calculation utility like it's done in *scope_EClass_feature(context, reference)* method in the {YourDSL}ScopeProvider class. As this scoping methods simply use the eType of the given reference the reimplementation should be straightforward.
Use this scope calculation utility in {YourDSL}ProposalProvider to propose values for the introduced EAttribute.
Optionally you can use this utility in a validation rule to add a warning/info to this EAttribute if it's not "valid".
Finally use the utility in your generator to create output based on valid target eObjects.
I also ran into this problem when creating a DSL to provide declerations of variables for a none-declerative language for a transition pahse. This method works but ask yourself if you realy want to have those nasty may-references.
You can drop the auto generated error in you UI module only. To do so, provide an ILinkingDiagnosticMessageProvider and override the function getUnresolvedProxyMessage:
class DSLLinkingDiagnosticMessageProvider extends LinkingDiagnosticMessageProvider {
override getUnresolvedProxyMessage(ILinkingDiagnosticContext context) {
if(context.context instanceof YourReference) {
// return null so the your error is left out
null
} else {
// use super implementation for others
super.getUnresolvedProxyMessage(context)
}
}
}
All linker-errors for YourReference will be missed. But be aware that there will be a dummy referenced object with all fealds null. Exspecialy the name ist lost and you can not set it due to a CyclicLinkingException. But you may create a new method that sets the name directly.
Note that the dummy object will have the type you entered in your gramma. If its abstract you can easily check witch reference is not linked.
I'm using code contract (actually, learning using this).
I'm facing something weird to me... I override a method, defined in a 3rd party assembly. I want to add a Contract.Require statement like this:
public class MyClass: MyParentClass
{
protected override void DoIt(MyParameter param)
{
Contract.Requires<ArgumentNullException>(param != null);
this.ExecuteMyTask(param.Something);
}
protected void ExecuteMyTask(MyParameter param)
{
Contract.Requires<ArgumentNullException>(param != null);
/* body of the method */
}
}
However, I'm getting warnings like this:
Warning 1 CodeContracts:
Method 'MyClass.DoIt(MyParameter)' overrides 'MyParentClass.DoIt(MyParameter))', thus cannot add Requires.
[edit] changed the code a bit to show alternatives issues [/edit]
If I remove the Contract.Requires in the DoIt method, I get another warning, telling me I have to provide unproven param != null
I don't understand this warning. What is the cause, and can I solve it?
You can't add extra requirements which your callers may not know about. It violates Liskov's Subtitution Principle. The point of polymorphism is that a caller should be able to treat a reference which actually refers to an instance of your derived class as if it refers to an instance of the base class.
Consider:
MyParentClass foo = GetParentClassFromSomewhere();
DoIt(null);
If that's statically determined to be valid, it's wrong for your derived class to hold up its hands and say "No! You're not meant to call DoIt with a null argument!" The aim of static analysis of contracts is that you can determine validity of calls, logic etc at compile-time... so no extra restrictions can be added at execution time, which is what happens here due to polymorphism.
A derived class can add guarantees about what it will do - what it will ensure - but it can't make any more demands from its callers for overridden methods.
I'd like to note that you can do what Jon suggested (this answers adds upon his) but also have your contract without violating LSP.
You can do so by replacing the override keyword with new.
The base remains the base; all you did is introduce another functionality (as the keywords literally suggest).
It's not ideal for static-checking because the safety could be easily casted away (cast to base-class first, then call the method) but that's a must because otherwise it would violate LSP and you do not want to do that obviously. Better than nothing though, I'd say.
In an ideal world you could also override the method and call the new one, but C# wouldn't let you do so because the methods would have the same signatures (even tho it would make perfect sense; that's the trade-off).