当前位置:首页 >> >>

Modularity and reusability in attribute grammars


Modularity and Reusability in Attribute Grammars
U. Kastens and W. M. Waite CU-CS-613-92 September 1992

University of Colorado at Boulder
Technical Report CU-CS-613-92 Department of Computer Science Campus Box 430 University of Colorado Boulder, Colorado 80309

Modularity and Reusability in Attribute Grammars
U. Kastens and W. M. Waite September 1992
Abstract
An attribute grammar is a declarative speci cation of dependence among computations carried out at the nodes of a tree. Attribute grammars have proven remarkably di cult to decompose into logical fragments. As a result, they have not yet been accepted as a viable speci cation technique. By combining the ideas of remote attribute access and inheritance, we have been able to de ne \attribution modules" that can be reused in a variety of applications. As an example, we show how to de ne reusable modules for name analysis that embody di erent scope rules.

1 Introduction
A fairly standard decomposition of the compiler construction task has evolved over the past twenty years, and most compilers have very similar structures. Many of the subproblems de ned by the standard decomposition can be described as computations over trees, in which information is attached to individual tree nodes and used to control various decisions.22 For example, consider the subproblem of determining the meaning of an operator that appears in an expression. The Pascal statement \A := B + C ;" might be represented internally by the tree fragment shown in Figure 1a. Here each node is classi ed as a Statement, V ariable, Expression, Operator or Identifier. Because there are di erent ways to construct nodes of each of these classes, the speci c rule used is given in parentheses below the node class. (The class can be deduced from the rule, so only an indication of the rule is physically stored in the tree.) According to the de nition of Pascal, the meaning of the the dyadic expression's operator might be integer addition, real addition or set union. In order to determine the meaning, the compiler might decorate the tree by attaching additional information to the nodes as shown in Figure 1b. The decoration AddReal would then be used in the decision to emit (say) a oating-point add instruction as the translation of the dyadic expression node. An algorithm for decorating a tree has three distinct components:

Computations The actual computations that result in individual decorations. Traversal The order in which the computations must be carried out at the various nodes
of the tree. Storage How intermediate results are to be stored during the decoration.

In Figure 1b, a computation determines that the operator is AddReal, given the fact that the indication is plus and the types of the children are both real. The traversal guarantees that this computation takes place after the computations of the indication and the types of the children. If the indication is not used in any other computation, it need not actually be stored in the tree node; the storage component of the algorithm might provide a single global variable that could hold all indications (because their lifetimes would not overlap). Under certain rather broad conditions, it is possible to mechanically derive both the traversal and the storage components of a decoration algorithm from the dependences among the computations.10,11 An attribute grammar is a speci cation of computations and dependence based on a formal calculus introduced by Knuth.16 By specifying only an attribute grammar, and then allowing some tool to derive the decoration algorithm, a designer avoids 2

Statement (Assignment)

Variable (Simple Variable)

Expression (Dyadic)

Identifier A

Expression (IdnExpr)

Operator (PlusOp)

Expression (IdnExpr)

Identifier B

Identifier C

a. A Pascal tree fragment
Expression (Dyadic) type=real

Expression (IdnExpr) type=real

Operator (PlusOp) operation=AddReal

Expression (IdnExpr) type=integer

b. The dyadic expression decorated with type information Figure 1: A Tree Decoration Problem

3

the need to think about tree traversal strategies and ways to store values used temporarily during decoration. The designer's task is thereby simpli ed. Despite this advantage, however, attribute grammars have not come into general use as a compiler speci cation method. We assert that the main reason for their lack of popularity has been the style in which they are normally written { a style that hinders both modular decomposition and reuse of speci cations. There are several di erent notations for describing attribute grammars.1 Section 2 reviews the formal calculus and illustrates the basic notation used to describe computations and dependence. Experience has shown that this notation must be augmented in order to be suitable for practical applications of attribute grammar speci cations. In Section 3 we discuss higher-level constructs that can be mapped to the basic constructs of the calculus, and in Section 4 we apply them to construct modules for a common tree decoration problem.

2 Basic Attibute Grammar Concepts
An attribute grammar is a quadruple AG = (G, A, R, B), where

G = (T; N; P; Z ) is a reduced context-free grammar, A = R = B =
X 2T N p2P p2P

A(X ) is a nite set of attributes

R(p) is a nite set of attribute computations B (p) is a nite set of plain computations

A(X ) \ A(Y ) 6= ; implies X = Y . For each occurrence of X in the derivation of a sentence of L(G), at most one attribute computation is applicable for the computation of each attribute a 2 A(X ). For each application of a production p in the derivation of a sentence of L(G), all computations of R(p) and B (p) are carried out.
The context-free grammar de nes the structure of the tree being decorated. Each production describes a context consisting of a node and its children (if any). (Figure 1b is an example of such a context.) Attribute values decorate the nodes, attribute computations specify how those values are related, and plain computations extract information for other processes. 4

SYMBOL Expr: type: DefTableKey; RULE Dyadic: Expr ::= Expr Op Expr COMPUTE Expr 1].type = SignatureElt(Op.operation,0); Op.operation = Identify2(Op.indication,Expr 2].type,Expr 3].type); END; SYMBOL Op: indication: Delimiter, operation: DefTableKey; RULE PlusOp: Op ::= '+' COMPUTE Op.indication = plus; IF(NotValid(Op.operation) message("Invalid operand types for this operator")); END;

Figure 2: Attribute Grammar Fragment Describing Figure 1
Figure 2 shows a fragment of an attribute grammar that describes the decorations of Figure 1b in a typical notation. SYMBOL constructs are used to specify the types of attributes in A(X ). Each RULE corresponds to one production of the reduced context-free grammar. Both the production name and the production itself are speci ed. (Note that the literal + does not actually appear in the tree; it is used only to aid the user in understanding the speci cation.) The elements of R(p) (the attribute computations) and B (p) (the plain computations) are given between the keywords COMPUTE and END. The order in which the computations are written down is totally irrelevant. Attribute computations describe relationships among attributes, not an algorithm for computing the values of those attributes. Our de nition is a slight generalization of those usually found in the literature, where the result of B (p) is restricted to Boolean values.10,23 It is important to emphasize that our generalization abstracts from the domains of attribute values and functions used in the computations, and thus models only the dependence among those computations: Each attribute e ectively represents a postcondition reached after completion of the attribute computation of R de ning that attribute. Attributes used in a computation of R or B e ectively represent preconditions for beginning that computation. If the dependence relationships satisfy certain constraints, it is possible to mechanically derive a complete computation algorithm from them.23 The derived algorithm is guaranteed to satisfy all pre- and postconditions. It is important to understand the rami cations of our view of the formal calculus, because it broadens the class of computations in formal speci cations.12,13 Most treatments of attribute grammars use attributes solely for the propagation of values. They demand 5

that functions appearing in computations of R and B have no side e ects, because such side e ects cannot be re ected in the attribute values. By considering attributes to represent pre- and postconditions, however, side e ects in component functions are easily accounted for.6 Of course this generalized view of the calculus does not exclude attribute grammar speci cation languages that require a strictly value passing model; it does, however, allow systematic and e cient implementations of many language processing tasks that are quite tedious for such models. We shall provide examples to support this contention in Section 4.

3 Paradigm Shifts
The two most complete attribute grammars that have appeared in the literature are those for Pascal25 and Ada.20 Both of these grammars were developed for the GAG system,25 a generator that accepts an ordered attribute grammar and produces a Pascal program to carry out attribute evaluation. The grammars are set in xed-width typewriter font of normal size, with a reasonable but not excessive amount of white space; Pascal requires 63 pages, Ada 395 pages. In each case, the authors have taken pains to structure the speci cation for the human reader. Since 1982, when the Pascal and Ada grammars were published, we have accumulated signi cant experience with attribute grammars as a speci cation mechanism. On the basis of our experience, we have developed a number of paradigms for notation and usage that simplify compiler descriptions. These paradigms have been employed successfully in a range of compilers built using Eli.5 The generalized view of the calculus underlying attribute grammars (described in Section 2) represents one fundamental paradigm shift; the others will be presented in this section. A complete example that shows how they enhance modular decomposition of an attribute grammar and provide reusable components is given in Section 4.

3.1 Remote Attribute Access
The formal calculus of Section 2 imposes the principle of locality in speci cations: All of the preconditions and the postcondition for a computation must appear within the context of a single node and its children. Locality of speci cations usually supports comprehensibility of attribute grammars, but sometimes one of the preconditions for a computation may be established by another computation that is far away in the tree. The calculus requires such a precondition to be established locally by computations provided in the intermediate 6

contexts. (This is the source of huge numbers of so called \copy rules" in attribute grammars formulated using only the notation of the underlying calculus.) In most cases, remote dependence follows one of three simple patterns: A computation at the root of a subtree containing the local context establishes the precondition. The precondition is the union of the postconditions for some set of computations at nodes that are descendants of the subtree rooted in the local context. The dependence involves an invariant for some iterative computation visiting nodes in (depth- rst) left-to-right order. Each of the three patterns can be directly formulated by an attribute grammar speci cation construct, as shown below. Those constructs can be transformed into the notation of the underlying calculus by generating all the intermediate computations mechanically, so they do not alter the theoretical properties of the attribute grammar. Direct formulation of remote dependence reduces the size of an attribute grammar, but an even more important characteristic is that it abstracts from the intermediate tree structure. It is invariant with respect to modi cations of that structure, a requirement for reuse of attribute grammar modules from libraries. In the remainder of this subsection we give a simple example of each of the three remote dependence patterns. As an example of a computation whose precondition is a computation at the root of a subtree containing the local context, consider the lexical nesting depth of a block. (The block nesting depth is used in forming addresses for variables declared within the block.23) Suppose that the nesting depth of a block is represented by an attribute Depth at the node corresponding to the block. The precondition for computation of the nesting depth is that a value be available for the nesting depth of the immediately enclosing block. If we assume that a block is one form of a statement, then this relationship might be expressed by the following attribute grammar rules:
RULE Program: Root ::= Block COMPUTE Block.Depth = 0; END; RULE Inner: Statement ::= Block COMPUTE Block.Depth = ADD(INCLUDING Block.Depth, 1); END;

7

Here the keyword INCLUDING indicates that the precondition is the rst instance of the speci ed attribute found when walking up the tree from the current node. Thus the rst argument of ADD is the nesting depth of the immediately enclosing block. If the single attribute name in this example is replaced by a parenthesized list, the precondition is the rst element of that list encountered when walking up the tree from the current node. A precondition that is the union of the postconditions for some set of computations at descendant nodes is illustrated by the problem of determining the cost of each statement in a program. Suppose that each operator has an associated cost (e.g. the length of the code generated for that operator), and that the cost of a statement is the sum of the costs of its component operators. If the cost of any operator or statement is represented by an attribute Cost at nodes corresponding to operators or statements, then the precondition for computation of the cost of a statement is that values be available for the costs of all of its component operators:
RULE Calculation: Statement ::= Expression COMPUTE Statement.Cost = CONSTITUENTS Operator.Cost WITH (int, ADD, IDENTICAL, ZERO); END;

Here the keyword CONSTITUENTS indicates that the precondition is the union of the instances of the given attribute appearing in the subtree rooted in the current node. Thus the cost of the statement depends on the union of the costs of the operators used in that statement. If the single attribute name in this example is replaced by a parenthesized list, the precondition is the union of all instances of all attributes on the list appearing in the subtree rooted in the current node. In this example, CONSTITUENTS must yield a value: the sum of the operator costs. The WITH clause speci es the type of the \union" value (int) and names three user-de ned functions used in its computation. The rst function (ADD) combines two \union" values to yield a \union" value, the second (IDENTICAL) creates a \union" value from an Operator.Cost value, and the third (ZERO) creates a \union" value from nothing. An invariant for a left-to-right iteration is illustrated by the computation of enumerated constant values in Pascal. Pascal enumeration constants are de ned by a list of identi ers, and the value represented by each identi er in the list is just the number of identi ers preceding it. Thus the rst identi er represents 0 because there are no identi ers to its left, the second represents 1, and so on. A user describes an invariant by declaring a special kind of attribute called a \chained attribute", or simply a \chain". The invariant is established initially by a CHAINSTART 8

directive at the root of some subtree. Computations updating the invariant may be associated with any node in that subtree. If Count represents the number of identi ers to the left of the current enumerated constant, the appropriate computations are expressed by the following rules:
CHAIN Count: int; RULE Enumeration: Type ::= '(' EnumConstList ')' COMPUTE CHAINSTART EnumConstList.Count = 0; END; RULE ConstListElt: EnumConst ::= Identifier COMPUTE EnumConst.Count = ADD(EnumConst.Count, 1); END;

Invariants can be expressed in terms of the underlying calculus by using two attributes per node. One is an inherited attribute representing the precondition for computations in the context of that node and its children, and the other is a synthesized attribute representing the postcondition established by those computations. Each nonterminal node in the subtree has such an attribute pair for each distinct invariant.

3.2 Symbol Computations
The block nesting depth computation discussed in Section 3.1 had two interesting properties: It only depended on attributes of a single symbol or on remote attributes. It should be applied at (almost) every instance of that symbol in the tree. When a computation has these properties, it can be associated with the symbol itself rather than with each of the di erent contexts in which the symbol appears. In Section 3.1, we assumed that a program was a block, and that a block was one kind of statement. One computation (setting the nesting depth to 0) was associated with the former context and another (incrementing the nesting depth) was associated with the latter. Suppose that, as in ALGOL 60, one possible implementation of a procedure body were also a block. This would result in two rules with identical computations:
RULE Inner: Statement ::= Block COMPUTE

9

Block.Depth = ADD(INCLUDING Block.Depth, 1); END; RULE Proc: Body ::= Block COMPUTE Block.Depth = ADD(INCLUDING Block.Depth, 1); END;

It would then pay to associate the nesting depth increment with the symbol Block instead of each context in which a block appears:
SYMBOL Block COMPUTE INH.Depth = ADD(INCLUDING Block.Depth, 1); END;

Here the keyword INH indicates that Depth is an inherited attribute and that its computation has to be associated with each context having Block on the right-hand side. The computation of a synthesized Block.i would be denoted SYNT.i and would be associated with each context having Block on the left-hand side. A computation that does not de ne an attribute is associated with each context having its symbol on the left-hand side. There can be several computations of either class in a single symbol attribution. The nesting depth increment must be replaced by a nesting depth initialization when the block is the entire program:
RULE Program: Root ::= Block COMPUTE Block.Depth = 0; END;

Here the computation given explicitly in the rule overrides the computation introduced implicitly by the use of Block on the right-hand side.

3.3 Inheritance
Several syntactic constructs of a language often share some set of semantic properties. For example, in ALGOL 68 there are three constructs that serve as scopes for identi er declarations: programs, serial clauses and procedures.24 To simplify the description of the semantics, ALGOL 68's designers introduced an additional symbol, \range", de ned in the grammar as representing all of these three constructs. \Range" could not be derived from 10

the axiom of the grammar, so there was never any phrase corresponding to that symbol in a program; it was used only in descriptions of the semantics to avoid having to make the same assertions about scope rules for identi ers in programs, serial clauses and procedures. The technique of inheritance allows one to obtain the same bene ts in an attribute grammar. The block nesting depth we have been discussing is used to access objects stored in a particular activation record at execution time. In ALGOL 60, distinct activation records may be associated with each procedure and each block. Thus procedures and blocks share the semantic property of being associated with distinct activation records at execution time. To simplify the description of this semantic property, we might introduce an additional symbol, Contour, representing it:
SYMBOL Contour COMPUTE INH.Depth = ADD(INCLUDING Contour.Depth, 1); END;

Note that this computation is now completely independent of the symbols used in a particular language de nition to denote constructs associated with distinct activation records. It abstracts the semantic concept of activation records addressed at execution time via a static chain.23 An abstract computation can be inherited by some set of symbols representing nodes in the tree:
SYMBOL Block INHERITS Contour END; SYMBOL Procedure INHERITS Contour END;

Each symbol may also inherit computations from several other symbols, possibly through several levels of inheritance. Inherited computations may be overridden by including explicit computations in SYMBOL declarations as well as by including them in RULE declarations.14

3.4 Cumulative Attribution
Cumulative attribution is a simple and e ective notational technique for decomposition of large attribute grammar speci cations: The user is allowed to write an arbitrary number of rules with the same production. Since each context in a tree is uniquely de ned by a single production in a reduced context-free grammar, the total set of computations for that context is the union of the sets of computations described by each rule with that production. 11

This makes it possible to decompose a speci cation into components, each covering one aspect of the total problem. For example, a compiler writer might provide separate descriptions of name analysis, overload resolution and translation into a target representation. Each of these tasks can be described by a few computations in a small number of contexts, using the technique of remote attribute access discussed above. The technique can also be applied to symbol computations, allowing an arbitrary number of speci cations describing di erent computations for the same symbol. A physical decomposition of an attribute grammar can be easily achieved by storing each component in a distinct le. The input to the attribute grammar processor is then a set of les, which cumulative attribution merges into a single speci cation in the underlying calculus. Some of those les might belong to libraries of reusable speci cation modules, each of which describes a single aspect of the problem to be solved. Examples of such speci cation modules are given in Section 4.

4 Attribution Modules for Scope Rules
Consistent renaming,23 the task of determining which source language entity is denoted by each identi er occurrence in a program, is de ned by the scope rules of the source language. In this section we show how the consistent renaming task can be embodied in scope-rule speci c attribution modules that use a scope-rule independent module15 for their computations. When developing an attribute grammar for a new language, an appropriate one of these attribution modules can easily be used to implement consistent renaming.

Consistent renaming is often merged with the solution of other compiler subproblems, such as that of converting symbol strings to an internal representation and that of associating properties with de ned entities. Such solutions cannot easily be reused in other compiler implementations. Our approach separates consistent renaming from other tasks by providing distinct modules for each. Source language entities are described by sets of properties (e.g. the lexical level of a variable, the value of a constant, or the amount of storage required for objects of a type). Each entity can be characterized by a unique key that allows the compiler to access these properties. The consistent renaming task therefore determines the appropriate key for each identi er occurrence in the program: It e ectively changes the programmer names (identi ers) to compiler names (keys). Access to the properties of an entity are provided by a di erent module, which is not concerned with relationships between identi ers and keys.

12

Computations for consistent renaming can be de ned in terms of three abstract data types:

Symbol The compiler's internal representation of an identi er. Environment The compiler's internal representation of a scope. DefTableKey The compiler's internal representation of a key for a program entity.
We assume that the Identifier-class leaves of the tree have been decorated with appropriate Symbol values when the tree representing a program is built, and therefore the computation of Symbol values is beyond the scope of this paper.23 The computations associated with the Environment abstract data type can be provided e ciently by a module that is independent of any speci c scope rules.15 Figure 3 de nes the part of the module interface that is relevant for this paper. No operations of the DefTableKey abstract data type are used explicitly during consistent renaming (the DefineIdn operation of the Environment abstract data type obtains DefTableKey values when needed). Therefore we will not consider details of the DefTableKey abstract data type in this paper.

4.1 ALGOL 60 Scope Rules
ALGOL 6019 was the rst language that de ned scope rules for de nitions in nested ranges. The basic concept can be summarized by the following rule: A de nition of an identi er a is visible in the smallest enclosing range with the exception of inner ranges that also contain a de nition of a. Hence, an applied occurrence of an identi er a identi es a de nition of a in the smallest enclosing range. De nitions of a in outer ranges are hidden. Figure 4 states this concept in terms of symbol attributes. Four symbols are used in Figure 4 to describe the syntactic constructs involved in the scope rules:

Range A region containing de nitions
13

NewEnv() Creates and returns a new Environment value representing an outermost scope. NewScope(e) Creates and returns a new Environment value representing a scope nested

within the scope represented by e. De neIdn(e,i) If the Symbol value i is not associated with a DefTableKey value in the scope represented by e, then a new DefTableKey value is obtained and associated with i in e. The DefTableKey value associated with i in e is returned. AddIdn(e,i,k) If the Symbol value i is not associated with a DefTableKey value in the scope represented by e, then the DefTableKey value k is associated with i in e and the Boolean value true is returned. Otherwise no associations are changed and the Boolean value false is returned. KeyInEnv(e,i) If the Symbol value i is associated with a DefTableKey value in the scope represented by e then that value is returned. Otherwise, if e represents the outermost scope then the distinguished DefTableKey value NoKey is returned. Otherwise the value of KeyInEnv(e',i), where e' is the Environment value representing the scope immediately enclosing e, is returned. KeyInScope(e,i) If the Symbol value i is associated with a DefTableKey value in the scope represented by e then that value is returned. Otherwise the distinguished DefTableKey value NoKey is returned.

Figure 3: Operations of the Standard Environment Module

14

SYMBOL Range: Env: Environment, GotLocalKeys, GotAllKeys: VOID; SYMBOL IdDef, IdUse: Sym: int, Key: DefTableKey; SYMBOL Root INHERITS Range COMPUTE INH.Env = NewEnv(); INH.GotAllKeys = THIS.GotLocalKeys; END; SYMBOL Range COMPUTE INH.Env = NewScope(INCLUDING Range.Env); INH.GotAllKeys = THIS.GotLocalKeys DEPENDS_ON INCLUDING Range.GotAllKeys; SYNT.GotLocalKeys = CONSTITUENTS IdDef.Key; END; SYMBOL IdDef COMPUTE SYNT.Key = DefineIdn(INCLUDING Range.Env, THIS.Sym); END; SYMBOL IdUse COMPUTE SYNT.Key = KeyInEnv(INCLUDING Range.Env, THIS.Sym) DEPENDS_ON INCLUDING Range.GotAllKeys; END;

Figure 4: An Attribution Module for ALGOL 60-like Scope Rules

15

Root The outermost region containing de nitions IdDef A de ning occurrence of an identi er IdUse An applied occurrence of an identi er
is the de nition table key associated with an identi er occurrence, and also represents the condition \this identi er occurrence is associated with a de nition table key". CONSTITUENTS IdDef.Key therefore represents the condition \all de ning occurrences in the subtree rooted here are associated with de nition table keys". Because the CONSTITUENTS construct computes a synthesized attribute of the symbol Range, however, the set over which the union is taken does not include any occurrences of IdDef lying inside subtrees that are themselves rooted in Range nodes.12 The attribute GotLocalKeys therefore represents the condition \all de ning occurrences in this range are associated with de nition table keys".
Key

\All de ning occurrences in the current range and all enclosing ranges are associated with de nition table keys" is a precondition for obtaining the key associated with an applied occurrence. The attribute GotAllKeys represents that precondition: Its computation combines \all de ning occurrences in enclosing ranges are associated with de nition table keys" with \all de ning occurrences in the current range are associated with de nition table keys". Figure 4 is completely independent of the abstract syntax of a particular language. In order to use it as a module for a particular language analysis, its symbols must be mapped to phrases of an abstract syntax for that language. For example, suppose that an abstract syntax for ALGOL 60 were to be developed from the grammar of the Revised Report.19 To use the module, we need distinct symbols corresponding to de ning and applied occurrences of identi ers. Unfortunately, the Revised Report does not make this distinction. Instead, all occurrences of identi ers in rules from Section 5 of the report are de ning occurrences, while occurrences of identi ers in rules from other sections of the report are applied occurrences. To solve this problem, de ne two new symbols for each symbol that represents an identi er in the original grammar. Replace all instances of the original symbol that represent de ning occurrences with one of the new symbols, and all instances of the original symbol that represent applied occurrences with the other. (For example, VarIdDef and VarIdUse might be de ned for the symbol variable_identifier.) The result of this modi cation is to move the speci cation of de ning and applied occurrences from the text of the report to the abstract grammar. The module of Figure 4 is linked to the modi ed abstract syntax by providing a \linkage" speci cation that resides in a distinct le:
SYMBOL VarIdDef INHERITS IdDef END;

16

SYMBOL VarIdUse INHERITS IdUse END; ...

Additional restructuring of the abstract grammar is needed to deal with the Range symbol. An ALGOL 60 block symbol represents a phrase that has exactly the semantics of a Range, but the procedure_declaration does not. Here is one of the productions for a procedure_declaration:
procedure_declaration: 'procedure' procedure_heading procedure_body

The problem is that the procedure_heading phrase contains both the procedure identi er (which is a de ning occurrence in the Range containing the procedure declaration) and the formal parameters (which are de ning occurrences in the Range constituting the procedure declaration). In order to place the identi er occurrences in their proper ranges, de ne a new symbol ProcRange to inherit the computations of a Range, and alter the abstract syntax to:
procedure_declaration: 'procedure' ProcIdDef ProcRange ProcRange: formal_parameter_part ';' value_part specification_part procedure_body

This formulation separates the procedure identi er from the formal parameters, placing them in distinct phrases. The proper association is then obtained by including two lines in the linkage speci cation:
SYMBOL ProcIdDef INHERITS IdDef END; SYMBOL ProcRange INHERITS Range END;

4.2 C Scope Rules
The basic scope rule concept of C deviates from that of ALGOL 60: A de nition of an identi er a is visible from its de nition point up to the end of the smallest enclosing range with the exception of inner ranges that also contain a de nition of a. To see the e ect of this rule, consider the following example: 17

{

int {

a = 1; int b = a; float a; a = b / 3;

} }

Here b is initialized with the value of the int variable a. The last statement assigns to the float variable a. ALGOL 60 scope rules would cause an unde ned result because b would be initialized to the (uninitialized) value of the float variable a. C scope rules are most easily implemented by an iterative process whose invariant is \all identi er occurrences preceding this point are associated with de nition table keys", as shown in Figure 5. This invariant is represented by the chain Env, which also provides access to the current environment. The CHAINSTART computation will be introduced into each context de ned by a production with Root or Range on the left-hand side. HEAD denotes the rst symbol on the right-hand side of that production. A new environment is established within each subtree rooted in a Root or Range symbol. In the case of a Range symbol, the new environment cannot be established within the subtree until the value of the Env environment is available at that Range node. Establishment of the new environment within the subtree does not a ect the value of the environment made available to nodes following the Range node in the iteration, but all possible identi er uses within that Range must be identi ed before any de nitions following the Range are entered into the environment. The invariant of the iteration must be re-established at each identi er occurrence. This is done by associating a de nition table key with the identi er occurrence, and making the invariant depend upon that key. De ning occurrences are distinguished from applied occurrences by the function invoked. The module of Figure 5 de nes the same symbols as the module of Figure 4, and interacts with the abstract syntax in the same way.

4.3 Modula-2 Scope Rules
Modula-2 follows ALGOL 60 scope rules for nested procedures (Figure 6), but also uses the concept of a module that may have import and export lists. A module is a phrase that may occur within a procedure declaration. Normally, none of the identi ers visible in the procedure body are visible in the module, and none of the identi ers declared in the module 18

CHAIN Env: Environment; SYMBOL IdDef, IdUse: Sym: int, Key: DefTableKey; SYMBOL Root COMPUTE CHAINSTART HEAD.Env = NewEnv(); END; SYMBOL Range COMPUTE CHAINSTART HEAD.Env = NewScope(THIS.Env); THIS.Env = THIS.Env DEPENDS_ON TAIL.Env; END; SYMBOL IdDef COMPUTE SYNT.Key = DefineIdn(THIS.Env, THIS.Sym); THIS.Env = THIS.Env DEPENDS_ON THIS.Key; END; SYMBOL IdUse COMPUTE SYNT.Key = KeyInEnv(THIS.Env, THIS.Sym); THIS.Env = THIS.Env DEPENDS_ON THIS.Key; END;

Figure 5: An Attribution Module for C-like Scope Rules

19

SYMBOL Range: Env: Environment, GotLocalKeys, GotAllKeys: VOID; SYMBOL IdDef, IdUse: Sym: int, Key: DefTableKey; SYMBOL Root INHERITS Range COMPUTE INH.Env = NewEnv(); INH.GotAllKeys = THIS.GotLocalKeys; END; SYMBOL Range COMPUTE INH.Env = NewScope(INCLUDING Range.Env); INH.GotAllKeys = THIS.GotLocalKeys DEPENDS_ON INCLUDING Range.GotAllKeys; SYNT.GotLocalKeys = CONSTITUENTS (IdDef.Key, Module.Exports); END; SYMBOL IdDef COMPUTE SYNT.Key = DefineIdn(INCLUDING Range.Env, THIS.Sym); END; SYMBOL IdUse COMPUTE SYNT.Key = KeyInEnv(INCLUDING Range.Env, THIS.Sym) DEPENDS_ON INCLUDING Range.GotAllKeys; END;

Figure 6: An Attribution Module for Modula-2 Ranges
are visible in the procedure body. If an identi er visible in the procedure body occurs on an import list in the module, however, that identi er becomes visible in the module. Similarly, if an identi er declared in the module occurs on an export list in the module, that identi er becomes visible in the procedure body. This last point gives rise to the only di erence between Figure 6 and Figure 4: The inclusion of symbols exported by local modules as locally-de ned symbols in the surrounding range. Figure 7 de nes the behavior of a module in Modula-2. By overriding the computation of the environment, Figure 7 makes identi ers from the surrounding environment invisible unless they are explicitly imported. (GotAllKeys, the condition \all de ning occurrences in the current range and all enclosing ranges are associated with de nition table keys", is also overridden to ensure that the necessary identi ers have been imported.) The computations for occurrences of identi ers on the import and export lists make de nitions in one scope visible in another. An exported identi er must be de ned in the 20

SYMBOL Module: Surround: Environment, Exports, GotAllSurroundingKeys: VOID; SYMBOL Module INHERITS Range COMPUTE INH.Env = NewScope(INCLUDING Root.Env); INH.GotAllKeys = THIS.GotLocalKeys DEPENDS_ON CONSTITUENTS ImportedId.Imported; INH.Surround = INCLUDING Range.Env; SYNT.Exports = CONSTITUENTS ExportedId.Exported; INH.GotAllSurroundingKeys = INCLUDING Range.GotAllKeys; END;

SYMBOL ExportedId: Exported, Sym: int, Key: DefTableKey; SYMBOL ImportedId: Imported, Sym: int, Key: DefTableKey; SYMBOL ExportedId COMPUTE SYNT.Key = KeyInScope(INCLUDING Range.Env, THIS.Sym) DEPENDS_ON INCLUDING Range.GotLocalKeys; SYNT.Exported = AddIdn(INCLUDING Module.Surround, THIS.Sym, THIS.Key); END; SYMBOL ImportedId COMPUTE SYNT.Key = KeyInEnv(INCLUDING Module.Surround, THIS.Sym) DEPENDS_ON INCLUDING Module.GotAllSurroundingKeys; SYNT.Imported = AddIdn(INCLUDING Range.Env, THIS.Sym, THIS.Key); END;

Figure 7: Modules, Export and Import in Modula-2

21

scope from which it is being exported, so KeyInScope is used to obtain the key. Once the key is available, AddIdn associates that key with the identi er in the surrounding environment. KeyInEnv is used to obtain the key for an imported identi er, since any identi er valid in the surrounding scope can be legally imported.

5 Related Work
The techniques described in this paper result from fteen years' experience using attribute grammars for compiler construction. All of them are implemented by LIGA,12 one of the tools integrated into the Eli compiler construction system.5 This section gives a brief summary of their origins, and compares them with other approaches to achieving modularity and reusability in attribute grammars. Our rst attribute grammar processor, GAG,25 was a direct implementation of the fundamental calculus described in Section 2. The idea of abstracting from the domain of attribute values and using attribute dependence to control side e ects of computations was introduced in 1982,21 and led to the initial design of LIGA in 1986. Remote attribute access using INCLUDING and CONSTITUENTS was rst introduced in GAG. The importance of chain attribution was recognized early, but special notation was used only in very restricted situations (e.g. the \bucket brigade"9). CHAIN and cumulative attribution were introduced into LIGA in 1990;12 symbol computations and inheritance appear for the rst time in this paper. We have emphasized that our approach supports decomposition of an attribute grammar into speci cations of the individual tasks to be carried out. Other approaches reported in the literature decompose the grammar along structural lines, independent of function. For example, in the decompositions based on object-oriented extensions 7,8,17,18 inheritance of attribution is tied to chain productions in the underlying concrete or abstract syntax. This means that computations can be reused only in contexts that are immediate descendants in the tree, or in which certain substructures coincide. Our mechanism, on the other hand, abstracts from the stucture so that computations can be reused with almost any abstract syntax. (Hedin's abstract productions8 have an e ect similar to our remote attribute access.) We advocate association of computations with symbols, but reuse can also be achieved by associating computations with \rule patterns".2 Unfortunately this approach does not support hierarchical organization of the speci cations, and it needs rather complicated overriding rules to describe the e ect of patterns which usually overlap. All the techniques mentioned so far map the speci cations to basic concepts of attribute grammars as our approach does. Any of the well-known implementation techniques for attribute evaluators can therefore be applied to generate code from them. Even though the 22

original speci cation is modular, the nal evaluator is monolithic. Another approach introduces a new evaluation model that decomposes the evaluator according to the speci cation structure.3 Reusability of such modules is restricted to grammars that match in their corresponding parts. The components interact at 1-visit boundaries, a hard restriction for the speci cation of language properties like scope rules that involve several subgrammars and often need more than one visit. Finally it should be mentioned that a modular decomposition of attribute grammars was suggested by attribute coupled grammars.4. That approach is aimed primarily at the speci cations of attributed tree transformation decomposed into sequences of transformation phases: Each module operates on a speci c tree computed by the preceding phase. Modules can thus only be reused for speci cations that have the same tree structure.

6 Conclusion
Attribute grammars provide a notation for specifying relationships among computations on a tree. The designer describes only individual computations, and a tool deduces both a traversal strategy and a strategy for storing computed values. Since the designer understands and maintains the speci cation, not the algorithm, the tool is free to use any implementation satisfying the given relationships. Logically distinct computations may be interleaved by the tool to minimize the number of traversals and the amount of temporary storage. Thus the tool may be able to signi cantly improve performance over an algorithm for the same computations that is understandable and maintainable. We have given an interpretation of the fundamental calculus on which attribute grammars are based that abstracts from the value domain of the computations, providing control of computations with side e ects. Using this interpretation, we have characterized patterns of dependence that allow us to abstract from the details of the tree structure as well. This abstraction permits us to construct modules de ning abstract semantics that are applicable to a variety of situations. We illustrated the process with speci cations of consistent renaming semantics for several languages. Descriptions like those in Section 4 are formal, yet easily understood. Proofs of their correctness can be constructed by available means,6 and they can be used to produce code without human intervention. Thus they can be used unchanged as the basis for both a language de nition and a compiler. This means that the compiler is guaranteed to implement the language de nition with respect to the feature described. The Eli system5 supports the techniques described in this paper, and has been used to construct compilers for a variety of languages. Tree decoration processes can also characterize 23

a number of problems in the processing of structured text and other applications. In each case, attribute grammars are useful in specifying the computation without having to provide details of traversal strategy and intermediate storage. Eli is capable of producing standalone modules that implement tree decoration algorithms for use as components of other systems. These modules are competitive in performance with hand-coded modules solving the same problems; if the decoration process has several logical parts, the generated modules are usually superior. By enabling modularity and reusability of speci cations, we belive that the techniques in this paper should overcome the disadvantages of classical attribute grammars, making their use as common as the use of context-free grammars. This should help to shorten the development time for new languages, and encourage the introduction of special problemoriented languages to simplify repetitive programming tasks.

Acknowledgements
This work was partially supported by the US Army Research O ce under grant DAAL03-92G-0158, and by the Ministry of Science and Technology of Nordrhein-Westfalen, Germany, under grant IVA3-10701691.

24

7 References
1. Deransart, P., Jourdan, M. & Lorho, B., Lect. Notes in Comp. Sci. #323, Springer Verlag,
New York{Heidelberg{Berlin, 1988.

2. Dueck, G. D. P. & Cormack, G. V., \Modular Attribute Grammars," The Computer Journal 33 (1990), 164{172. 3. Farrow, R., Marlowe, T. J. & Yellin, D. M., \Composable Attribute Grammars: Support
for Modularity in Translator Design and Implementation," Proc. of ACM Symposium on Principles of Programming Languages (1992).

4. Ganzinger, H. & Giegerich, R., \Attribute Coupled Grammars," SIGPLAN Notices 19 (June 1984), 157{170. 5. Gray, R. W., Heuring, V. P., Levi, S. P., Sloane, A. M. & Waite, W. M., \Eli: A Complete, Flexible Compiler Construction System," Communications of the ACM 35 (Feb. 1992),
121{131.

6. Gries, D., The Science of Programming, Springer Verlag, 1981. 7. Grosch, J., \Ag - An Attribute Evaluator Generator," GMD Forschungsstelle Karlsruhe,
Germany, Report 16, 1989.

8. Hedin, G., \An Object-Oriented Notation for Attribute Grammars," Proc. of the European Conf. on Object-Oriented Programming, ECOOP'89 (1989).

9. Jullig, R. K. & DeRemer, F., \Regular Right-Part Attribute Grammars," SIGPLAN Notices 19 (June 1984), 171{178. 10. Kastens, U., \Ordered Attribute Grammars," Acta Informatica 13 (1980), 229{256. 11. Kastens, U., \Implementation of Visit-Oriented Attribute Evaluators," in Proceedings
of the International Summer School on Attribute Grammars, Application and Systems, Lect. Notes in Comp. Sci. #545, Springer Verlag, New York{Heidelberg{Berlin, 1991, 114{139.

25

12. Kastens, U., \Attribute Grammars as a Speci cation Method," in Proceedings of the International Summer School on Attribute Grammars, Application and Systems, Lect. Notes in Comp. Sci. #545, Springer Verlag, New York{Heidelberg{Berlin, 1991, 16{47. 13. Kastens, U., \Attribute Grammars in a Compiler Construction Environment," in Proceedings of the International Summer School on Attribute Grammars, Application and Systems, Lect. Notes in Comp. Sci. #545, Springer Verlag, New York{Heidelberg{Berlin, 1991, 380{400. of the LIGA System, 1992.

14. Kastens, U., \LIDO - Short Reference," University of Paderborn, FRG, Documentation 15. Kastens, U. & Waite, W. M., \An Abstract Data Type for Name Analysis," Acta Informatica 28 (1991), 539{558. 16. Knuth, D. E., \Semantics of Context-Free Languages," Mathematical Systems Theory 2 (June 1968), 127{146. 17. Koskimies, K., \Object-Orientation in Attribute Grammars," University of Tampere, Finland, A-1991-1, April 1991.

18. Magnusson, B., Bengtsson, M., Dahlin, L. O., Fries, G., Gustavsson, A., Hedin, G., Mi-

nor, S., Oscarsson, D. & Taube, M., \An Overview of the Mjolner/Orm Environment: Incremental Language and Software Development," Lund University, Report LU-CSTR:90:57, 1990.

19. \Revised Report on the Algorithmic Language ALGOL 60," Communications of the ACM 6 (Jan. 1963), 1{17. 20. Uhl, J., Drossopoulou, S., Persch, G., Goos, G., Dausmann, M., Winterstein, G. & Kirchgassner, W., An Attribute Grammar for the Semantic Analysis of ADA, Lect. Notes in Comp. Sci. #139, Springer Verlag, Berlin, 1982. 21. Waite, W. M., \Generator for Attribute Grammars - Abstract Data Type," Gesellschaft
fur Mathematik und Datenverarbeitung, Arbeitspapier 219, Karlsruhe, BRD, Sept. 1986.

22. Waite, W. M., \Use of Attribute Grammars in Compiler Construction," in Attribute Grammars and their Applications, Pierre Deransart & Martin Jourdan, eds., Lect. Notes in Comp. Sci. #41, Springer Verlag, Berlin, 1990, 255{265.
26

23. Waite, W. M. & Goos, G., Compiler Construction, Springer Verlag, New York, NY, 1984. 24. Wijngaarden, A. van, Mailloux, B. J., Peck, J. E. L. & Koster, C. H. A., \Report on the Algorithmic Language ALGOL 68 ," Numerische Mathematik 14 (1969 ), 79{218 . 25. Zimmermann, E., Kastens, U. & Hutt, B., GAG: A Practical Compiler Generator, Lect. Notes in Comp. Sci. #141, Springer Verlag, Heidelberg, 1982.

27



相关文章:
更多相关标签: