1. Sumant Tambe, et. al LEESA DSPD 2008 An Embedded Declarative Language for Hierarchical Object Structure Traversal Sumant Tambe* Aniruddha Gokhale Vanderbilt University, Nashville, TN, USA *Contact : sutambe@dre.vanderbilt.edusutambe@dre.vanderbilt.edu 2nd International Workshop on Domain-Specific Program Development (DSPD), GPCE 2008, Nashville, TN, USA

2. 2 Sumant Tambe, et. al LEESA DSPD 2008 Object Structures and Operations  Object Structures  Hierarchical data structure (XML)  Object Graph/Network (e.g., Domain-specific Models)  Uses of object structures  Domain-specific modeling language (DSML) interpreters, transformations  Systems using XML data binding  Component deployment & configuration engines  Web services  Common operations  Queries (search)  Traversals (visit each element)

3. 3 Sumant Tambe, et. al LEESA DSPD 2008 Existing Approaches of Accessing Object Structures S1 S2  Imperative object-oriented API  Implemented using classes, interfaces, and collections of objects in Java, C++  Visitor pattern: separates visitation actions from traversal (Good!)  Fancy C++ generic programming techniques using STL/Boost::bind  XQuery / XSLT / XPath  W3C standards  Suitable for hierarchical structures  Suitable for XML-to-XML transformation  Traversal/Visitor Language (TVL)  Separate specification language  Autogenerates traversal code  LINQ (Language Integrated Query)  Used in C# and VB  Based on Microsoft.NET framework TraversalVisitation Action class StateMachine { … }; class State { … }; class Transition { … };

4. 4 Sumant Tambe, et. al LEESA DSPD 2008 Limitations of Existing Techniques  Imperative object-oriented API  Code bloat while using large/complex data models  Intent is lost in low level details  In classic visitor, “visit-all” semantics are inefficient in most cases (visits more elements than necessary)  In other visitor variants traversals are coupled with visitation actions – limits reusability  XQuery / XSLT / Xpath  Only hierarchical. No support for object graphs  Traversal/Visitor Language (TVL)  High cost of development of lexer/parser  Must learn new language (syntax, semantics) and code generator  LINQ  Depends on language extensions in C#, VB  Depends on.NET framework features Visitation Action Traversal &

5. 5 Sumant Tambe, et. al LEESA DSPD 2008 Solution: LEESA  LEESA: Language for Embedded Query and Traversal  Features of LEESA  A domain-specific embedded language (DSEL) for writing queries and traversals over object graphs (models)  Provides short and expressive syntax for traversal  Supports multiple traversal strategies  Supports Visitor: decouples visitation actions from traversal  Cheap to develop because it is embedded in C++  Reuses C++ lexer, parser and whole slew of standard libraries  Much flatter learning curve than external DSLs  Seamless transition between LEESA and traditional C++  Interoperable with other popular C++ DSELs (e.g., Boost.Phoenix, Boost.Lambda)

6. 6 Sumant Tambe, et. al LEESA DSPD 2008 Layered Architecture of LEESA DSL for query and traversal A C++ generative programming technique OO access API (UDM, XML data binding) In memory representation of object structure

7. 7 Sumant Tambe, et. al LEESA DSPD 2008 Composition Navigation Using LEESA Example 1: Give all the States under RootFolder RootFolder() >> StateMachine() >> State() Example 2: Give all the Property atoms under RootFolder Returns a std::set RootFolder() >> StateMachine() >> State() >> Property() Returns a std::set Example 3: Give the parent State of Property P P << State() Returns a std::set of size = 1 Parent/child composition relationship navigation using “>>” and “<<”

8. 8 Sumant Tambe, et. al LEESA DSPD 2008 Object Structure Queries Using LEESA Example 4: Give all the States under RootFolder whose name starts with “XYZ” RootFolder() >> StateMachine() >> State() >> SelectByName(State(),”XYZ.*”) Example 5: Same as above + apply arbitrary filter + sort RootFolder() >> StateMachine() >> State() >> SelectByName(State(),”XYZ.*”) >> Property() >> Select(Property(), predicate) >> Sort(Property(), comparator) Many query operators are supported such as Unique, Sort, SelectByName, Select (generic) Where predicate and comparator are standard C++ functions or stateful function objects

9. 9 Sumant Tambe, et. al LEESA DSPD 2008 Association Navigation Using LEESA Association ends can be navigated to using “>>&” followed by the name of the association. Example 6: Give all the States that are at the “dst” end of transitions RootFolder() >> StateMachine() >> Transition() >>& Transition::dstTransition Returns a std::set = { S2, S3, S4 }

10. 10 Sumant Tambe, et. al LEESA DSPD 2008  Visitor pattern is supported  Add visitor object after “>>”  Calls corresponding Visit function  Separates visitation actions from traversal order Object Structure Visitation Using LEESA CountVisitor cv; RootFolder() >> StateMachine() >> cv >> State() >> cv >> Property() >> cv  “>>” visits in breadth-first order  “>>=” visits in depth-first order  Query operators work with both strategies  Traversal strategies can be mixed together RootFolder() >>= StateMachine() >> cv >>= State() >> cv >>= Property() >> cv Example 7: Visit each StateMachine, State, and Property and count them Example 8: Same as before but in depth-first order

11. 11 Sumant Tambe, et. al LEESA DSPD 2008 Combining LEESA Traversal Strategies #define DEPTH_FIRST >>= CountVisitor pv; RootFolder() DEPTH_FIRST StateMachine() >> cv >> State() >> Property() >> Sort(Property(),comparator) >> cv Example 9: Visit all StateMachines in depth-first order and visit all the properties in sorted order

12. 12 Sumant Tambe, et. al LEESA DSPD 2008 Layered Architecture of LEESA DSL for query and traversal A C++ generative programming technique OO access API (UDM, XML data binding) In memory representation of object structure

13. 13 Sumant Tambe, et. al LEESA DSPD 2008 LEESA’s Generic Data Access Layer class RootFolder { set StateMachine_kind_children(); template set children (); }; class StateMachine { set State_kind_children(); set Transition_kind_children(); template set children (); }; class State { set Property_kind_children(); template set children (); }; class Transition { State srcTransition(); State dstTransition(); }; class Property; Automatically generated UDM C++ interface from meta-model: T determines children kind

14. 14 Sumant Tambe, et. al LEESA DSPD 2008 Layered Architecture of LEESA DSL for query and traversal A C++ generative programming technique OO access API (UDM, XML data binding) In memory representation of object structure

15. 15 Sumant Tambe, et. al LEESA DSPD 2008 Overview of Expression Templates  First developed in 1994 by Todd Veldhuizen  Known uses: Boost.Spirit parser, Blitz++ scientific computing library  Based on the idea of lazy evaluation  Execute expressions much later in the program from the point of their definition  Pass an expression -- not the result of the expression -- as a parameter to a function  E.g., foo (x * y + z);  The result of expression is computed before calling the function  Instead, an abstract syntax tree (AST) is created at compile-time using templates and operator overloading XY Z + *

16. 16 Sumant Tambe, et. al LEESA DSPD 2008 Returns a set of R LEESA’s Expression Templates  LEESA defines a couple of expression templates and many operators  We’ll consider GetChildren operator  ChainExpr expression template  ChainExpr creates an AST that embodies traversal GetChildren  L.children (); RootFolder() >> StateMachine() >> State() >> Property() GetChildren  StateMachine.children ();

17. 17 Sumant Tambe, et. al LEESA DSPD 2008 Future Work & Concluding Remarks  How to improve error reporting?  Promising solution: C++ concept checking  How to make it less dependent on structure?  Structure-shyness property  How to support wildcards? RootFolder() >> ***** >> Property()