1. Honors Compilers Semantic Analysis and Attribute Grammars Mar 5th 2002

2. Semantic analysis Enforce context-dependent language rules that are not reflected in the BNF, e.g.a function must have a return statement. Enforce context-dependent language rules that are not reflected in the BNF, e.g.a function must have a return statement. Decorate AST with semantic information for subsequent code generation, e.g. determine types of all expressions Decorate AST with semantic information for subsequent code generation, e.g. determine types of all expressions Expand complex constructs in preparation for code generation, e.g. aggregates into loops. Expand complex constructs in preparation for code generation, e.g. aggregates into loops. General framework: compute attributes General framework: compute attributes

3. Attributes and attribute grammars Syntax-directed framework: Syntax-directed framework: For every symbol in the grammar we define some computable properties (e.g. the value of a constant expression) For every symbol in the grammar we define some computable properties (e.g. the value of a constant expression) For every production in the grammar we give computation rules for the properties of all symbols on both sides of the production (e.g. the value of a sum is the sum of the values of the operands) For every production in the grammar we give computation rules for the properties of all symbols on both sides of the production (e.g. the value of a sum is the sum of the values of the operands) The rule is local: it only refers to other symbols in the same production The rule is local: it only refers to other symbols in the same production The evaluation of the attributes can require an arbitrary number of traversals of the AST: arbitrary context dependence (.e.g. the value of a declared constant is found in the constant declaration) The evaluation of the attributes can require an arbitrary number of traversals of the AST: arbitrary context dependence (.e.g. the value of a declared constant is found in the constant declaration)

4. Example: binary strings String ::= Digit | String Digit String ::= Digit | String Digit Digit ::= 0 | 1 Digit ::= 0 | 1 Attribute: the numeric value of a string Attribute: the numeric value of a string Digit ::= ‘0’ Digit ::= ‘0’ Val digit := 0; Val digit := 0; Digit ::= ‘1’ Digit ::= ‘1’ Val digit := 1; Val digit := 1; String ::= Digit String ::= Digit Val String ::= Val digit Val String ::= Val digit String ::= String Digit String ::= String Digit Val String1 := 2 * Val String2 + Val digit Val String1 := 2 * Val String2 + Val digit

5. Example : type imposed by context Assignment ::= Name := Expression Attribute : if expression is overloaded, its type is determined by the type of the name: Context_Type expression := Type name Two distinct attributes: type of name is determined from context and/or other rules

6. Inherited and synthesized attributes If attribute of left-hand side is computed from attributes in the right-hand side, attribute is synthesized: bottom-up propagation If attribute of left-hand side is computed from attributes in the right-hand side, attribute is synthesized: bottom-up propagation If attribute of symbol on right-hand is computed from attributes of left-hand side, or from attributes of other symbols on right-hand side, attribute is inherited: top-down propagation of information If attribute of symbol on right-hand is computed from attributes of left-hand side, or from attributes of other symbols on right-hand side, attribute is inherited: top-down propagation of information StringVal NameContext typeexpression

7. General results Attribute grammars have the power of Turing machines Attribute grammars have the power of Turing machines Attributes are computed by repeated passes over the AST Attributes are computed by repeated passes over the AST Attribute definitions may be cyclic; checking whether an attribute grammar has cycles is decidable but potentially expensive Attribute definitions may be cyclic; checking whether an attribute grammar has cycles is decidable but potentially expensive In practice inherited attributes are handled by means of global data structures (symbol table) In practice inherited attributes are handled by means of global data structures (symbol table) Useful subsets: L-attributed and S-attributed grammars Useful subsets: L-attributed and S-attributed grammars

8. L-attributed grammars and S-attributed grammars In an L-attributed grammar, inherited attributes can be computed left-to-right: In an L-attributed grammar, inherited attributes can be computed left-to-right: N ::= S1 S2 S3 N ::= S1 S2 S3 attributes of S1 cannot depend on attributes of S2 or S3 attributes of S1 cannot depend on attributes of S2 or S3 Easy to implement in a top-down parser: when building a node, nodes on which its attributes depend have been seen and processed. Easy to implement in a top-down parser: when building a node, nodes on which its attributes depend have been seen and processed. An S-attributed grammar has only synthesized attributes An S-attributed grammar has only synthesized attributes Usable with bottom-up parser Usable with bottom-up parser Can convert L- to S- but result is awkward. Can convert L- to S- but result is awkward.

9. Some important attributes For expressions: type For expressions: type For overloaded calls: candidate interpretations For overloaded calls: candidate interpretations For identifiers: entity (defining_occurrence) For identifiers: entity (defining_occurrence) For definitions: scope For definitions: scope For data/function members: visibility (public, protected, private) For data/function members: visibility (public, protected, private) For function: virtual functions (primitive operations) For function: virtual functions (primitive operations) Etc, etc. Etc, etc.

10. Attribute computation and tree traversals In gnat (and others) mostly left-to-right, top-down traversal, with localized multiple traversals In gnat (and others) mostly left-to-right, top-down traversal, with localized multiple traversals Inherited attributes computed during declaration processing, symbol table carries inherited information as one global data structure Inherited attributes computed during declaration processing, symbol table carries inherited information as one global data structure Synthesized attributes on terminals: names, literal values Synthesized attributes on terminals: names, literal values In the presence of overloading, type is both inherited and synthesized: two passes required over expressions. In the presence of overloading, type is both inherited and synthesized: two passes required over expressions. Generated code can be treated as synthesized attribute Generated code can be treated as synthesized attribute

11. Name Resolution Compute attribute entity: associate every identifier (use occurrence) with the corresponding defining occurrence. Compute attribute entity: associate every identifier (use occurrence) with the corresponding defining occurrence. If entity is overloaded, associate entity with set of candidate entities, to be resolved by types and context. If entity is overloaded, associate entity with set of candidate entities, to be resolved by types and context. Complications: Complications: Block structure and hiding rules Block structure and hiding rules Context and import rules Context and import rules

12. Name resolution and block structure Basic rule: inner definition hides outer one with same name Basic rule: inner definition hides outer one with same name Data structures reflect scope nesting Data structures reflect scope nesting A stack of scopes: defining occurrences of functions, packages, blocks, loops, records A stack of scopes: defining occurrences of functions, packages, blocks, loops, records A list of local entities declared in each scope A list of local entities declared in each scope A names table A names table Entry in names table points to innermost occurrence of entity with given name Entry in names table points to innermost occurrence of entity with given name All identifiers with given name point to same names table entry (handled by scanner) All identifiers with given name point to same names table entry (handled by scanner) Name resolution does not require any hashing Name resolution does not require any hashing

13. Data structures for name resolution Entity chain, homonym chain, chars Entity chain, homonym chain, chars Outer scope var1county Parent scope var1Z2yzz Current scope var1county1 var1county var1 Names table

14. Semantic actions for visibility processing A scope is any entity that has local declarations: package, function, record type, task, block. A scope is any entity that has local declarations: package, function, record type, task, block. On scope entry: place new scope on stack, initialize list of local entities On scope entry: place new scope on stack, initialize list of local entities For every declaration: chain name entry to local entity, set homonym of local entity to outer entity with same name For every declaration: chain name entry to local entity, set homonym of local entity to outer entity with same name On scope exit: chain name entry to homonym of local entity. Local entity becomes invisible. On scope exit: chain name entry to homonym of local entity. Local entity becomes invisible. On package exit: leave entries chained, because entities remain visible through qualified notation. On package exit: leave entries chained, because entities remain visible through qualified notation. Full information remains in the tree for subsequent passes. Full information remains in the tree for subsequent passes.

15. Resolving qualified names To resolve A.B, first resolve A (direct name). To resolve A.B, first resolve A (direct name). If A is enclosing scope, follow homonym chain for B until we find a variable whose scope is A If A is enclosing scope, follow homonym chain for B until we find a variable whose scope is A If A is a variable, find its type If A is a variable, find its type If record or struct, find component of type named B If record or struct, find component of type named B If pointer, apply rule to designated type (implicit dereference) If pointer, apply rule to designated type (implicit dereference) If task, find entry named B If task, find entry named B To resolve A.B.C, recurse: resolve prefix A.B, then apply previous rules To resolve A.B.C, recurse: resolve prefix A.B, then apply previous rules To resolve A->B (C++): type of A must be of the form *T, proceed as above To resolve A->B (C++): type of A must be of the form *T, proceed as above

16. Packages and import rules Packages (namespaces) are scopes, have chain of local entities, but are not in scope stack. Packages (namespaces) are scopes, have chain of local entities, but are not in scope stack. Package entities also linked to names table Package entities also linked to names table To resolve P.A, if P is a package in context (with_clause) use same algorithm as for an enclosing scope To resolve P.A, if P is a package in context (with_clause) use same algorithm as for an enclosing scope If use_clause is active for P, then A is a legal use of P.A Treat as a direct name if there is no A in an enclosing scope that hides it: entities in enclosing scopes hide imported entities. If use_clause is active for P, then A is a legal use of P.A Treat as a direct name if there is no A in an enclosing scope that hides it: entities in enclosing scopes hide imported entities. If multiple entities are use-visible, reference is ambiguous If multiple entities are use-visible, reference is ambiguous

17. Detailed algorithms in gnat sources Visibility : sem_ch8 Visibility : sem_ch8 New_Scope: push current scope on scope_stack New_Scope: push current scope on scope_stack Find_Direct_Name: resolve identifier Find_Direct_Name: resolve identifier Find_Selected_Component: resolve A.B, where A can be an expression Find_Selected_Component: resolve A.B, where A can be an expression Find_Qualified_Name: resolve A.B, where A is a scope Find_Qualified_Name: resolve A.B, where A is a scope Use_One_Package: make entities that are local to a package use-visible in current context Use_One_Package: make entities that are local to a package use-visible in current context End_Use: when use clause goes out of scope, the converse End_Use: when use clause goes out of scope, the converse End_Scope: remove local entities from visibility, pop scope_stack End_Scope: remove local entities from visibility, pop scope_stack Declarations: sem_ch3, sem_util Declarations: sem_ch3, sem_util Enter_Name: enter new entity into local chain, diagnose duplicates Enter_Name: enter new entity into local chain, diagnose duplicates

18. Top-down processing: all but expressions Semantic analysis of package declaration: Semantic analysis of package declaration: Enter new scope Enter new scope Process visible declarations Process visible declarations Process private declarations Process private declarations Exit scope Exit scope Semantic analysis of while statement Semantic analysis of while statement Process (analyze and resolve) condition Process (analyze and resolve) condition Process list of statements Process list of statements Semantic analysis of object declaration Semantic analysis of object declaration Enter new entity into current scope Enter new entity into current scope Resolve type definition Resolve type definition Analyze and resolve expression Analyze and resolve expression