Attributes

Synthesized attributes

syn T A.x();

x is a synthesized attribute in class A and of type T.
There must be equations defining x in A (if A is concrete) or in all concrete subclasses of A (if A is abstract).
Note! Synthesized attributes are conceptually equivalent to abstract virtual functions (without side-effects). The main difference is that their values may be cached (see below). They can be accessed in the same way as virtual functions. I.e., the declaration generates the following Java API:

T A.x();

eq A.x() = Java-expression;

The equation defines the value of the synthesized attribute x of AST nodes of type A.
The Java-expression that defines the value must be free from externally visible side-effects. The context of the expression is the class A, and any part of the class A's API may be used in the computation, including accesses to other attributes.
Note! Equations defining synthesized attributes are conceptually equivalent to virtual method implementations.

eq B.x() = Java-expression;

Suppose B is a subclass to A. This equation overrides the corresponding (default) equation for A.x().
Note! This is equivalent to overriding method implementations.

refine S eq B.x() = Java-expression;

Equations defined in one aspect can be refined in another aspect, in the same way as methods can be refined, see JastAdd aspect files. In this example, the equation replaces the corresponding equation declared in the aspect S. The value from the original equation in S can be accessed by the expression S.B.x()

(This feature is available in JastAdd version R20051107 and later.)

Examples of refine eq on Google Code Search.

Inherited attributes

inh T A.y();

y is an inherited attribute in class A and of type T. There must be equations defining y in all classes that have children of type A. If a class has several children of type A, there must be one equation for each of them.

Inherited attributes can be accessed in the same way as synthesized attributes. I.e., the declaration generates the following Java API:

T A.y();

Note! Inherited attributes differ from ordinary virtual functions in that their definitions (equations/method implementations) are located in the parent AST node, rather than in the node itself.

Note! The concept of inherited attributes in this Attribute Grammar sense is completely different from object-oriented inheritance. Both attribute grammars and object-orientation were invented in the late 60's and the use of the same term "inheritance" is probably a mere coincidence: In AGs, inheritance takes place between nodes in a syntax tree. In OO, inheritance takes place between classes in a class hierarchy.

eq C.getA().y() = Java-expression;

This equation defines value of the inherited attribute y() of the A child of a C node.
Note! The Java-expression executes in the context of C.
Note! The equation is similar to a method implementation.
Note! The equation actually applies to all inherited attributes in the subtree rooted at A, provided that they declare the attribute. See below under broadcast attributes.

eq D.getA().y() = Java-expression;

Suppose D is a subclass of C. In this case, the equation overrides the previous one.
Note! This is analogous to overriding a virtual method implementation.

Shorthand for synthesized attributes

syn T A.x() = Java-expression;

The declaration of a synthesized attribute and the (default) equation for it can be written in one clause. So the clause to the left is equivalent to:

syn T A.x();
eq A.x() = Java-expression;

Method body

syn T A.x() {
   ...
   return Java-expression;
}

It is possible to write the computation of an attribute value as a method body instead of as a single expression. This may be convenient when the computation is complex. Inside the method body it is possible to use ordinary imperative Java code with local variables, assignments, loops, etc. However, the net result of the computation must not have any side-effects.

Lazy attributes

An attribute can be declared lazy in order to speed up the evaluation. An attribute that is declared lazy will automatically have its value is cached after the first access to it. The next time the attribute is accessed, the cached value is returned directly. We recommend that attributes that are expensive to compute and that are accessed multiple times should be declared lazy. For example, declaration bindings and type attributes are good candidates for caching. Future versions of JastAdd might include heuristics for automatically selecting which attributes to cache.

syn lazy A.x();

Here, the attribute x of class A is declared lazy.

Parameterized attributes

Attributes can have parameters. This is a bit unusual for attribute grammars, but a natural generalization when you view attributes as virtual functions.

syn T A.x(int a);
eq A.x(int a) {
   return Java-expression;
}

Here, x is a parameterized synthesized attribute. The equation is similar to a method implementation and the argument values can be used to compute the resulting value.

inh T A.y(int a);
eq C.getA().y(int a) {
   return Java-expression;
}

Here, y is a parameterized inherited attribute. The equation executes in the context of C and can in addition access the arguments (a in this case).

Broadcasting inherited attributes

Often, an inherited attribute is used in a number of places in a subtree. If basic inherited attributes are used, the value needs to be copied explicitly using inherited attributes in all the intermediate nodes. For convenience, JastAdd supports another technique, namely broadcasting of an inherited attribute to a complete subtree. An equation defining an inherited attribute actually broadcasts the value to the complete subtree of the child. By using this technique, no explicit copy attributes are needed.

eq C.getA().y() = ...;
inh T A.y();

Here, the equation defines an inherited attribute y() declared in the A child of a C node. This equation actually applies not only to the inherited y() attribute of the A child, but to all inherited y() attributes in the whole subtree of A. In order to for a node N in the subtree to access y(), the attribute must, however, be exposed by declaring y() as an inherited attribute of N.

inh T B.y();

Here, the attribute y() is exposed in B by declaring it as an inherited attribute there. If there is a B node that is in the subtree rooted at the A that is a child of a C node, then the equation above will apply.

Overruling broadcast definitions

A broadcast definition of an attribute a() applies to all nodes in a subtree rooted by N. If, however, there is a node in the subtree which has another equation that defines a() for a child M, that equation will take precedence for defining a() in M and its subtree.

Differentiating between children in a list

When defining an inherited attribute of a child node that is an element of a list, it is sometimes useful to know which index the child node has. This can be done as follows:

C ::= E*;
eq C.getE(int index).y() = ...index...

Here, a C node has a list of E children. When defining the y() attribute of a given (subtree of an) E child, the value might depend on the index of the child. For example, if the E nodes are actual arguments of a procedure, we might want to pass down the expected type of each argument.

The example equation shows how to declare the index as a parameter of the getE() method, and to access the index in the equation body.

Unconditional rewrite rule

rewrite A {
   to B {
      ...
      return exp;
   }
}

An A node will be replaced by the node specified in the Java expression exp. This will happen as soon as the A node is accessed (by a get() method from its parent), so if you traverse the tree you will only be able to access the final rewritten nodes.

A and B must be AST classes.

The exp must be of type B.

Let the set S be the superclasses of A (including A) that occur on right-hand sides of productions in the abstract syntax. B must be a subclass of all classes in S. This garantuees that replacing an A node by a B node does not break the rules in the abstract syntax.

The code in the body of the rewrite may access and rearrange the nodes in the subtree rooted at A, but not any other nodes in the AST. Furthermore, the code may not have any other side effects.

Conditional rewrite rule

rewrite A {
   when ( condition )
   to B {
      ...
      return exp 
   }
}

The conditional rewrite works in the same way as the unconditional one, but performs the replacement only if the boolean expression condition is true. The condition may access anything in the AST, e.g., attributes, other tree nodes and their attributes, etc.

Examples of rewrite on Google Code Search.

Iterative rewriting

After a node has been replaced according to a rewrite rule, all conditional rewrite rules are checked again, and a new rewrite may be performed. This is iterated until no rule conditions hold.

Order of rewriting

At each iteration of rewriting, the rule conditions are evaluated in a certain order. The first condition that is true is used for rewriting in that iteration. The order in which rule condition evaluation occurs is the following:

• conditions in superclasses are evaluated before conditions in subclasses
• conditions within an aspect file are evaluated in lexical order
• conditions in different aspect files are evaluated in the order the files are listed in the jastadd command.

Confluency

If the order of rewriting of a node does not effect the final result, the rules are said to be confluent. This is highly desirable, since it makes the specification more readable to not have to take lexical order of rules into account. However, JastAdd cannot check that the rules are confluent. In cases where several conditions for a node are true at the same time, we recommend that you contemplate the rules and try to find out if they could be non-confluent. In that case, we recommend you to refine the conditions so that only one can apply at a time. This makes your specification independent of lexical order. Note that it is often useful to have several different rules that apply at the same time for a given node, but which are confluent.

Shorthand notation

If you have several conditional rewrite rules, you may write them together. So, e.g., writing

rewrite A {
   when ( condition-1 )
   to B {
      ...
      return exp-1 
   }
   when ( condition-2 )
   to C {
      ...
      return exp-2 
   }
}

... is equivalent to:

rewrite A {
   when condition-1 
   to B {
      ...
      return exp-1 
   }
}
rewrite A {
   when condition-2 
   to C {
      ...
      return exp-2 
   }
}

Sometimes you don't need a block for computing the resulting node. It may be sufficient with an expression. In that case, you may simply write the expression instead of the block, e.g., as follows:

rewrite A {
   when ( condition-1 )
   to B exp-1 
   when ( condition-2 )
   to C exp-2 

}

... which is equivalent to

rewrite A {
   when ( condition-1 )
   to B { return exp-1 }
   when ( condition-2 )
   to C { return exp-2 }
}

Circular attributes

Attributes can be circularly defined. I.e., the value of the attribute can depend (indirectly) on itself. Circular attributes are evaluated iteratively, starting with a start value given in the declaration of the attribute. The evaluation stops when the value equals that for the previous iteration.
Circular attributes are always cached. They do not need to be declared "lazy".
If a lazy attribute is circular, but not declared as such, this will be detected at runtime, and an exception will be generated.
To be sure that the evaluation of circular attributes will converge, the values should be arranged into lattices of finite height, the bottom values should be used as starting values, and each equation on the cycle should be monotonic with respect to the lattices.

syn T A.x(int a) circular [bv];
eq A.x(int a) = rv;

Here, the attribute x is a circular attribute. The starting value is bv (a Java expression).
The equation defines x as having the value computed by the Java expression rv. Note that  rv may depend (directly or indirectly) on x.

Examples of syn circular at Google Code Search.

Examples of inh circular at Google Code Search (no matches 2009-July-09)

Newer syntax

syn nta C A.anNTA() = new C();

• NTAs can be inherited or synthesized.
• The value in the equation should be a freshly built AST subtree. It should be complete in the sense that all its children should also be freshly created nodes (i.e., they are not allowed to be initialized to null).
• The NTA can itself have attributes that can be accessed like normal attributes.
• If the NTA has inherited attributes, there must be equations for those attributes in some ancestor, as for normal children.
• Examples of syn nta at Google Code Search.
• Examples of inh nta at Google Code Search. (No matches on 2009-July-09)

Older syntax

In the older syntax, you introduce a nonterminal attribute as follows:

• Declare the NTA in the ast file, (See also NTAs in the abstract syntax).
• Declare the NTA as an attribute in a jrag file. It can be declared as a synthesized or an inherited attribute. The name of the attribute should be the same as in the AST traversal API, e.g., getX if the NTA is called X.
• Add equations defining the NTA. The defining value should be a new AST of the appropriate type, created using the AST creation API.

Note that if the NTA is a List or an Optional node, you need to create the appropriate AST with a List or an Opt node as its root. See examples below.

Simple synthesized NTA

A ::= B /C/;

syn C A.getC() = new C();

The NTA C is declared in the .ast file. It is then declared as a synthesized attribute getC() in the .jrag file. The equation is provided directly in the declaration and creates a new C node.

List NTA

A ::= B /C*/;

syn C A.getCList() =
   new List().
      add(new C()).
      add(new C());

The list NTA C* is declared in the .ast file. It is then declared as a synthesized attribute getCList() (the same name as in the implementation level traversal API) in the .jrag file. The equation is provided directly in the declaration and creates a List node to which is added a number of C nodes (two in this example).

Collection attributes

Collection attribute declaration

coll T A.c() [fresh] with m;

• Within square brackets, a Java expression, fresh, should be written that provides a fresh object of type T. For example, new T().
• The composite value will be built by the underlying JastAdd machinery by calling the method m for each contribution.
• The method m, should be a one-argument method of T.
• The method m should mutate the T object by adding a contribution to it.
• The method m should be commutative, in the sense that the order of calling m for different contributions should yield the same resulting T value. 
• Examples of coll declarations at Google Code Search.

Contribution declaration

N1 contributes value-exp
when cond-exp
to N2.a()
for N2-ref-exp;

• N2-ref-exp should be a Java expression denoting an object of type N2.
• The value-exp should be a Java expression of the argument type of the method m.
• The contribution is only applied when the boolean condition cond-exp is true.
• The "when"-clause may be omitted.

Contributions to a set of collections

N1 contributes value-exp
to N2.a()
for each N2-ref-set-exp;

• N2-ref-set-exp should denote an object should denote an object of the Java type java.lang.Iterable, which should contain objects of type N2.
• The contribution is added to each of these N2 objects.
• Examples of ccontribution declarations at Google Code Search.

Scoped collections

coll N1.a [fresh] with m root R

• If R is the same nodetype as N1 (or a subtype of N1), then the subtree root will be the N1 object.
• If R is some other nodetype, the subtree root will be the closets object of type R on the way from N1 towards the AST root.