Beliebte Suchanfragen

Cloud Native



Agile Methoden



Macro annotations in Scala 2

28.3.2023 | 11 minutes of reading time

In this blog post we will take a look at macro annotations, a powerful tool for code transformation and generation in Scala. Macro annotations allow us to transform the code of a definition, e.g., a class or method, at compile time. This can be used to reduce boilerplate code, because the compiler generates it for us. During compilation, a macro annotation will expand the abstract syntax tree (AST) of the annotated definition. The transformation code is written in Scala itself, making use of Scala's reflection API to analyze and transform the AST. Hence, the use of macros is also referred to as metaprogramming.

Macro annotations are one of two types of macros in Scala 2. The other one are def macros, which are explained in [1]. Macro annotations are an experimental feature and have to be enabled via compiler flag -Ymacro-annotations in Scala 2.13, or by using the macro paradise plugin in older version starting from Scala 2.10.

Caching method return values with macro annotations

As a simple example for macros in Scala 2 (version 2.13.10), we will implement a macro annotation for caching return values of arbitrary methods. The cache will store key-value pairs of the input parameters of the method as well as its return value and directly return them for cached inputs. The code we will see in the following can be found here.

We start by defining a Cache trait with a simple MapCache implementation:

1trait Cache[K, V] {
2  def put(key: K, value: V): Option[V]
3  def get(key: K): Option[V]
6class MapCache[K, V] extends Cache[K, V] {
7  private val map = mutable.Map.empty[K, V]
9  override def put(key: K, value: V): Option[V] = map.put(key, value)
10  override def get(key: K): Option[V] = map.get(key)

Before introducing the macro annotation, we first want to have a look at how an implementation without macros could look like. The caching logic can be generalized into a method with type parameters K and V for the key and value types:

1def cached[K, V](cache: Cache[K, V], input: K)(f: => V): V = {
2  cache.get(input) match {
3    case Some(value) =>
4      value
5    case None =>
6      val result = f
7      cache.put(input, result)
8      result
9  }

The method takes the cache and the input as parameters and wraps a function f. It looks up the input in the cache and returns the respective value if one is found. In case of a cache miss, the function f is executed and the result is cached and then returned.

This cached method can now be used in methods of arbitrary signature:

1val fCache = new MapCache[(Int, Int), Int]
3def f(x: Int, y: Int): Int = cached(fCache, (x, y)) {
4  x * y
7val gCache = new MapCache[(Int, String), String]
9def g(x: Int, s: String): String = cached(gCache, (x, s)) {
10  x.toString + s

The generic cached method provides some simplification for implementing the methods whose results we want to cache. However, we still have to define a cache for each method and pass it along with the input parameters to the cached method. This feels like boilerplate, which we can get rid of with a macro annotation.

Generating the caching logic

We introduce a cached macro annotation, which can be used to annotate a method of arbitrary signature, providing all the required caching logic. In the first step, we only implement the transformation of the annotated method itself. For now, the cache that is used by the annotated method still has to be defined in the application code. This will be improved in a later step.

1class cached extends StaticAnnotation {
2  def macroTransform(annottees: Any*): Any = macro cached.impl
5object cached {
6  def impl(c: whitebox.Context)(annottees: c.Tree*): c.Tree = ???

The cached annotation extends the StaticAnnotation class and defines a macroTransform method. The actual code transformation happens in the impl method, which gets the annottees along with a context as inputs.

The annottees are a sequence of ASTs, namely the definitions that are annotated by the cached annotation. In our example, there will always be one annottee, as we will annotate methods. Multiple annottees can occur, for instance, when a class with a companion object is annotated. In this case, both the class and its companion object are passed to the macroTransform method and can be transformed in its implementation.

The context provides access to Scala's reflection API, including features such as quasiquotes and term names, which we will make use of in the transformation code. We can import those directly as import c.universe._.

Now, let's have a look at the implementation of the macro transformation. This may seem overwhelming at first, but we will give a step-by-step explanation below:

1def impl(c: whitebox.Context)(annottees: c.Tree*): c.Tree = {
2  import c.universe._
4  annottees.head match {
5    case q"""$mods def $method[..$typeParams](
6      ...$params): $returnType = $rhs""" =>                        // (1)
8      val paramNames = {
9        case q"$_ val $name: $_ = $_" => name                      // (2)
10      }
12      val cacheName = TermName(s"${method}Cache")                  // (3)
13      val keyName = TermName(c.freshName("key"))
14      val resultName = TermName(c.freshName("result"))
16      val newRhs =                                                 // (4)
17        q"""
18         val $keyName = (..$paramNames)
19         $cacheName.get($keyName) match {
20           case Some(value) =>
21             value
22           case None =>
23             val $resultName = $rhs
24             $cacheName.put($keyName, $resultName)
25             $resultName
26         }
27        """
28      val expandedMethod = q"""$mods def $method[..$typeParams](
29        ...$params): $returnType = $newRhs"""                      // (5)
31      expandedMethod
33    case annottee => 
34      c.abort(annottee.pos, "Annottee must be a method")           // (6)
35  }

We start in (1) by pattern matching the annottee against a quasiquote, a code snippet wrapped inside an interpolated string q"...". Quasiquotes are a useful notation for constructing ASTs, which would otherwise have to be written as a bunch of nested method applications. As seen in the code, quasiquotes can also be used for pattern matching, allowing us to extract parts of the AST. Conceptually, this is similar to normal string interpolation with s"...", which allows inserting or extracting parts of the string. Note that in quasiquotes, we extract parts of the AST instead of strings. For instance, $mods are the modifiers of the annotated method (e.g. private) and rhs is the expression on the right-hand side of the annottee. Also note how the parameters and type parameters of the method are extracted. The ..$typeParams notation indicates a list of trees (in this case type trees). Meanwhile, the ...$params notation indicates a nested list of lists of trees, representing the parameter lists of the annotated method.

The pattern matching ensures, that the annottee is indeed a method. Otherwise, we jump to the default case (6), where we abort compilation with a corresponding error message.

In (2), we extract the names from the parameters, which we will use in the transformed right-hand side of the method to build the key that is looked up and stored in the cache. Then, we define some variable names in (3), which we will also use in the right-hand side. The $cacheName references a variable outside the method scope, which we assume to be present. If it is not, the compilation will fail. The other two variable names will reference local variables in the method body. To avoid any naming conflicts with existing variables, we create them calling the freshName method from Scala's reflection API, which ensures that the generated name will be unique.

The main part of the code transformation happens in (4), where we define newRhs, the new body of the method. The new right-hand side is more or less identical to the implementation of the cached method that we defined earlier. The key is created as a tuple built from the method parameters, using the (..paramNames) notation. In case of a cache miss, the original method body rhs is used to retrieve a result.

Finally, in (5), the expandedMethod is built, using all the parts of the original method, but replacing the right-hand side with newRhs. This method definition will replace the annotated method during compile time.

With this in place, we can annotate arbitrary methods as @cached. The compiler will then transform each method according to the macro transformation defined above. The application code now looks as follows, still requiring the explicit definition of the fCache and gCache for the two methods f and g, but already the method definitions themselves are simplified:

1val fCache = new MapCache[(Int, Int), Int]
4def f(x: Int, y: Int): Int = x * y
6val gCache = new MapCache[(Int, String), String]
9def g(x: Int, s: String): String = x.toString + s

Generating caches at compile time

Now that we have simplified the method body by implementing the cached macro annotation, we also want to get rid of the explicit Cache instance definitions. In order to achieve that, we need to not only transform a definition, but generate an entirely new definition as well, namely the cache. Luckily, macro annotations allow us to expand an annottee into as many ASTs as we want. This allows us to expand the annotated method into both the transformed method and an additional value definition for the cache. The impl method of the macro annotation is adapted as follows:

1def impl(c: whitebox.Context)(annottees: c.Tree*): c.Tree = {
2  import c.universe._
4  annottees.head match {
5    case q"""$mods def $method[..$typeParams](
6      ...$params): $returnType = $rhs""" =>
8      val (paramNames, paramTypes) = {
9        case q"$_ val $name: $tpt = $_" => name -> tpt             // (1)
10      }.unzip
12      val cacheName =                                              // (2)
13        TermName(c.freshName(method + "_generatedCache"))
14      val keyName = TermName(c.freshName("key"))
15      val resultName = TermName(c.freshName("result"))
17      val newRhs = ...
18      val expandedMethod = q"""$mods def $method[..$typeParams](
19        ...$params): $returnType = $newRhs"""
21      val cacheKeyType = tq"(..$paramTypes)"                       // (3)
22      val cache =                                                  // (4)
23        q"""
24         private val $cacheName: Cache[$cacheKeyType, $returnType] =    
25           new MapCache[$cacheKeyType, $returnType]
26         """
28      q"$expandedMethod; $cache"                                   // (5)
30    case annottee => 
31      c.abort(annottee.pos, "Annottee must be a method")
32  }

In (1) we extract the types of the method parameters, which we then use in (3) to define the key type of the cache. The interpolator tq"(..$paramTypes)" is used to build a tuple type of the paramTypes. The tq"..." notation is similar to q"...", which we learned previously, with the difference that the latter is used to build expression trees, while the former is used for type trees.

Instead of referencing a cache definition from application code, our new cacheName in (2) will reference a generated value definition. We use a fresh name for the cache, which avoids naming conflicts and also allows us to annotate multiple overloaded methods that share the same name but with different method signatures.

The cache definition is generated in (4). It uses the unique cacheName as name, the tuple of the method parameter types as key type and the return type as value type. We use the MapCache implementation as before.

Finally, in (5) we return both the expandedMethod and the cache as two individual definitions created from a single annottee.

With the caches generated from macro transformation as well, our application code reduces to only the methods with the @cached annotation. Any trace of actual caching logic is removed and deferred to code generation at compile time.

2def f(x: Int, y: Int): Int = x * y
5def g(x: Int, s: String): String = x.toString + s

Custom cache implementations from implicit cache factories

So far we have used a simple MapCache implementation for the caches. In a real-world example, one would rather use a more advanced caching solution, which provides features such as a size limit and expiration of entries. For example, we could use a cache implementation based on Guava, that would implement our previously defined Cache trait:

1class GuavaCache[K, V] extends Cache[K, V] {
3  private val cache = CacheBuilder.newBuilder()
4    .maximumSize(1000)
5    .expireAfterAccess(5, TimeUnit.MINUTES)
6    .build[K, V]
8  override def put(key: K, value: V): Option[V] = {
9    val oldValue = get(key)
10    cache.put(key, value)
11    oldValue
12  }
14  override def get(key: K): Option[V] = Option(cache.getIfPresent(key))

We would like to be able to use different Cache implementations with the @cached annotation without having to touch the macro transformation for each one. An easy way to achieve that, is by resolving the cache implementation from the implicit scope. First, we provide a CacheFactory that offers an apply method producing a Cache:

1trait CacheFactory {
2  def apply[K, V](): Cache[K, V]

An example implementation using the previously defined GuavaCache would look as follows. It is defined as implicit value in order to allow the generated code to resolve it:

1implicit val guavaCacheFactory: CacheFactory = new CacheFactory {
2  override def apply[K, V](): Cache[K, V] = new GuavaCache[K, V]

In the generated code, we replace the explicit creation of a MapCache with a call of Scala's implicitly method, which receives the CacheFactory as type parameter. The apply method is called on the resolved CacheFactory, yielding a Cache of the appropriate key and value type:

1val cache =
2  q"""
3   private val $cacheName: Cache[$cacheKeyType, $returnType] = 
4     implicitly[CacheFactory].apply[$cacheKeyType, $returnType]()
5  """

During implicit resolution, the compiler will look for an implicit definition that fits the required CacheFactory type and thereby find the guavaCacheFactory.


Using the caching example, we have seen how macro annotations can be used in Scala 2 to generate code at compile time. The cached annotation can be reused and is independent of the signature of the method it is applied to. Also, we bypass the overhead of defining a cache for every method whose return values we want to cache, by generating it as a new definition for every annotated method.

While macro annotations provide an interesting way to use metaprogramming in Scala, they are also somewhat clumsy to use in practice, especially with regard to IDE support, which is not always working properly. Taking IntelliJ IDEA using the Scala plugin as an example, the IDE is often unable to infer the type of calls to the reflection API properly. Quasiquotes are also somewhat difficult to work with, as they are essentially just strings containing Scala code, which are not type-checked in any form by the IDE plugin.

With the rework of the language in Scala 3, macro annotations have unfortunately been removed from the language. Only recently they have been added again as an experimental feature in the pre-release version 3.3.0-RC2. Compared to Scala 2, their usage is still rather clumsy. We will have a look at macro annotations in Scala 3 in the next blog post.



share post




More articles in this subject area

Discover exciting further topics and let the codecentric world inspire you.


Gemeinsam bessere Projekte umsetzen.

Wir helfen deinem Unternehmen.

Du stehst vor einer großen IT-Herausforderung? Wir sorgen für eine maßgeschneiderte Unterstützung. Informiere dich jetzt.

Hilf uns, noch besser zu werden.

Wir sind immer auf der Suche nach neuen Talenten. Auch für dich ist die passende Stelle dabei.