PMD's Java module has an extensive framework for the calculation of metrics, which allows rule developers to implement and use new code metrics very simply. Most of the functionality of this framework is abstracted in such a way that any PMD supported language can implement such a framework without too much trouble. Here's how.
Edit me

Internal architecture of the metrics framework

Overview of the Java framework

The framework has several subsystems, the two most easily identifiable being:

  • A project memoizer (ProjectMemoizer). When a metric is computed, it’s stored back in this structure and can be reused later. This reduces the overhead on the calculation of e.g. aggregate results (ResultOption calculations). The contents of this data structure are indexed with fully qualified names (JavaQualifiedName), which must identify unambiguously classes and methods.

  • The façade. The static end-user façade (JavaMetrics) is backed by an instance of a JavaMetricsFaçade. This allows us to abstract the functionality of the façade into pmd-core for other frameworks to use. The façade instance contains a project memoizer for the analysed project, and a metrics computer (JavaMetricsComputer). It’s this last object which really computes the metric and stores back its result in the project mirror, while the façade only handles parameters.

Metrics (Metric<N>) plug in to this static system and only provide behaviour that’s executed by the metrics computer. Internally, metric keys (MetricKey<N>) are parameterized with their version (MetricVersion) to index memoisation maps (see ParameterizedMetricKey<N>). This allows us to memoise several versions of the same metric without conflict.

At the very least, a metrics framework has those two components and is just a convenient way to compute and memoize metrics on a single file. The expressive power of metrics can be improved by implementing signature matching capabilities, which allows a metric to count signatures matching a specific pattern (a mask) over a whole class. This was originally designed to work across files, given a working usage resolution. However, making that work with incremental analysis is harder than it looks, and has been rescheduled to another project.

Abstraction layer

As you may have seen, most of the functionality of the first two components are abstracted into pmd-core. This allows us to implement new metrics frameworks quite quickly. These abstract components are parameterized by the node types of the class and operation AST nodes. Moreover, it makes the external behaviour of the framework very stable across languages, yet each component can easily be customized by adding methods or overriding existing ones.

The signature matching aspect is framed by generic interfaces, but it can’t really be abstracted more than that. The info given in the signatures is usually very language specific, as it includes info about e.g. visibility modifiers. So more work is required to implement that, but it can already be used to implement sophisticated metrics, that already give access to detection strategies.

Implementation of a new framework

1. Groundwork

  • Create a class implementing QualifiedName. This implementation must be tailored to the target language so that it can indentify unambiguously any class and operation in the analysed project. You must implement equals, hashCode and toString. Example
  • Determine the AST nodes that correspond to class and method declaration in your language. These types are referred hereafter as T and O, respectively. Both these types must implement the interface QualifiableNode, which means they must expose a getQualifiedName method to give access to their qualified name.

2. Implement the façade

  • Create a class extending AbstractMetricsComputer<T, O>. This object will be responsible for calculating metrics given a memoizer, a node and info about the metric. Typically, this object is stateless so you might as well make it a singleton.
  • Create a class extending BasicProjectMemoizer<T, O>. There’s no abstract functionality to implement. Example Example
  • Create a class extending AbstractMetricsFacade<T, O>. This class needs a reference to your ProjectMemoizer and your MetricsComputer. It backs the real end user façade, and handles user provided parameters before delegating to your MetricsComputer. Example
  • Create the static façade of your framework. This one has an instance of your MetricsFaçade object and delegates static methods to that instance. Example
  • Create classes AbstractOperationMetric and AbstractClassMetric. These must implement Metric<T> and Metric<O>, respectively. They typically provide defaults for the supports method of each metric. Example
  • Create enums ClassMetricKey and OperationMetricKey. These must implement MetricKey<T> and MetricKey<O>. The enums list all available metric keys for your language. Example
  • Create metrics by extending your base classes, reference them in your enums, and you can start using them with your façade!

Optional: Signature matching

You can match the signature of anything: method, field, class, package… It depends on what’s useful for you. Suppose you want to be able to match signatures for nodes of type N. What you have to do then is the following:

  • Create a class implementing the interface Signature<N>. Signatures describe basic information about the node, which typically includes most of the modifiers they declare (eg visibility, abstract or virtual, etc.). It’s up to you to define the right level of detail, depending on the accuracy of the pattern matching required.
  • Make type N implement SignedNode<N>. This makes the node capable of giving its signature. Factory methods to build a Signature<N> from a N are a good idea.
  • Create signature masks. A mask is an object that matches some signatures based on their features. For example, with the Java framework, you can build a JavaOperationSigMask that matches all method signatures with visibility public. A sigmask implements SigMask<S>, where S is the type of signature your mask handles.
  • Create utility methods in your abstract class metric class to count signatures matching a specific mask. Example