1. Department of Computer Science, Graduate School of Information Science & Technology, Osaka University Program Slicing on Java byte-code for Locating Functional Concerns Takashi Ishio † Ryusuke Niitani † Gail Murphy ‡ Katsuro Inoue † † Osaka University, Japan ‡ University of British Columbia, Canada

2. Department of Computer Science, Graduate School of Information Science & Technology, Osaka University Concern Location A functional concern is code that helps fulfill a functional requirement.  A software maintenance task usually focuses on a functional concern. Concern location comprises “Search and Explore.”  Search “interesting” methods grep or other feature location tools  Explore the interaction among the methods call graph, class hierarchy tree, cross reference

3. Department of Computer Science, Graduate School of Information Science & Technology, Osaka University Example: Autosave function in jEdit jEdit periodically saves the contents of text area.  A user can specify the frequency. We can easily find Autosave class, Buffer.autosave() method and BufferIORequest.autosave() method. How the classes and methods are interacting?

4. Department of Computer Science, Graduate School of Information Science & Technology, Osaka University Exploring Interaction among methods Important information: control-flow and data-flow.  Which method triggers the autosave function.  Which class has a necessary data (e.g. filename).  How a method saves the contents to a text file. We have to read following classes: Autosave, Buffer, BufferIORequest, PerspeciveManager, VFSManager, FileVFS …

5. Department of Computer Science, Graduate School of Information Science & Technology, Osaka University Automated Concern Location We are trying to extract a concern graph from code fragments specified by a developer.  Our approach is based on program slicing.  Our tool is based on Soot, a Java bytecode analysis framework. Program Slicing with Heuristics Slice-to-ConcernGraph Translation Code fragments related to a functionality A concern graph a program slice

6. Department of Computer Science, Graduate School of Information Science & Technology, Osaka University Autosave concern graph Input = Autosave.*(), Buffer.autosave(), BufferIORequest.autosave()

7. Department of Computer Science, Graduate School of Information Science & Technology, Osaka University Program Slicing Slicing extracts statements related to criteria statements specified by a user. 1. A program P is converted to a program dependence graph. vertices: statements in P edges: control/data dependence relations 2. A user specifies “slicing criteria” statements in P. The statements are translated into “criteria vertices” in the PDG. 3. A program slice, a set of statements that affect or depend on criteria, is extracted by graph traversal from criteria vertices. 1i = 3; 2if (a > 0) { 3 print i; 4} data dependence control dependence use definition

8. Department of Computer Science, Graduate School of Information Science & Technology, Osaka University Slice including unrelated concerns Slicing usually extracts many statements.  A functional unit is connected to other units by control/data-flow.  28% on average in C program † † Binkley, D., Gold, N. and Harman, M.: An Empirical Study of Static Program Slice Size. ACM TOSEM Vol.16, No.2, Article 8, April 2007. AutosaveUndoManager autosave_dirty flag activateset/reset reset CompleteWord set slicing

9. Department of Computer Science, Graduate School of Information Science & Technology, Osaka University Slicing with Barriers A barrier is a vertex or an edge that terminate graph traversal †. † Krinke, J.: Slicing, Chopping, and Path Conditions with Barriers. Software Quality Journal, Vol.12, No.4, pp.339-360,December 2004. AutosaveUndoManager autosave_dirty flag activateset/reset reset CompleteWord set A barrier blocks graph traversal. slicing

10. Department of Computer Science, Graduate School of Information Science & Technology, Osaka University Similarity-based Barrier The key idea is following: if two methods are contributing to the same functionality, the methods use similar methods, fields and classes. Name Set NS(m) = a set of types, classes, methods and fields referred in m. A long name is “tokenized”. e.g. “java.io.File”  “java”, “io”, “File”, “java.io.File”

11. Department of Computer Science, Graduate School of Information Science & Technology, Osaka University Example of Similarity package org.gjt.sp.util; class IntegerArray { private int[] array; private int len; public void add(int num){ if(len >= array.length) { int[] arrayN = new int[len * 2]; System.arraycopy(array,0,arrayN,0,len); array = arrayN; } array[len++] = num; } public final int getSize() { return len; } public final void setSize(int len) { this.len = len; } org.gjt.sp.util.IntegerArray, org, gjt, sp, util, integer, array, void, add, int, len, int[], java.lang.System, java, lang, system, arraycopy NS(IntegerArray.add) NS(IntegerArray.getSize) NS(IntegerArray.setSize) org.gjt.sp.util.IntegerArray, org, gjt, sp, util, integer, array, getSize, get, size, int, len org.gjt.sp.util.IntegerArray, org, gjt, sp, util, integer, array, len, setSize, set, size, void, int sim = 0.801 sim = 0.639

12. Department of Computer Science, Graduate School of Information Science & Technology, Osaka University Identifying Barriers Program slicing is blocked at method m if m is not related to slicing criteria Similarity(m, C) ≦ threshold  A method m is related to slicing criteria if slicing criteria includes a method n such that m is similar to n. C = a set of methods that contain slicing criteria vertices.

13. Department of Computer Science, Graduate School of Information Science & Technology, Osaka University Slicing algorithm Slicing with summary edges and barriers  defined by Horwitz  extended by Krinke PDG based on Jimple code  “jimple” is an intermediate representation for bytecode. 3-address code Simple control-flow: “if” and “goto” Independent of JVM stack operation Calculate similarity for each method Code fragments related to a functionality a program slice Identify barriers Slicing with Barriers

14. Department of Computer Science, Graduate School of Information Science & Technology, Osaka University Visualizing a slice as a concern graph Concern Graph  A vertex is a class, a method or a field.  An edge represents a relation between two vertices. call, create, check, read, write, superclass, … We applied rule-based translation. † v1 in m1 v2 in m2 m1 m2 call call or parameter SliceConcern Graph † Kameda, D. and Takimoto, M.: Building Cocnern Graph Based on Program Slicing. IPSJ Transactions on Programming, Vol.46, No.11 (Pro 26), pp.45-56. in Japanese. v1 in m1 READ obj.field m1 field read

15. Department of Computer Science, Graduate School of Information Science & Technology, Osaka University A graphical output with Graphviz We omit intra-class edges in graphical format.  Detail is provided in textual format. e.g. “Autosave.setInterval(interval) calls new Timer(interval, Autosave).”

16. Department of Computer Science, Graduate School of Information Science & Technology, Osaka University The effectiveness of barriers Barriers reduced concern graph size:  1000 methods  20 methods  Printable on an A3 or A4-sized paper Comparing extracted graphs with hand-made concern graphs (not finished yet). Our previous experiment is reported in: 仁井谷竜介，石尾隆，井上克郎 : プログラムスライシングを用いた機能的 関心事の抽出手法の提案と実装. PPL 2007. in Japanese. concern graph size on 6 maintenance tasks on jEdit and our Slicer

17. Department of Computer Science, Graduate School of Information Science & Technology, Osaka University Information extracted from Java program To construct a dependence graph  Control dependence relation  Data dependence relation  Call Graph (with dynamic binding information) To identify barriers  a set of types, methods, fields referred in each method m To slice the dependence graph  Mapping source code to vertices

18. Department of Computer Science, Graduate School of Information Science & Technology, Osaka University Slicing Tool Overview Java Class Files PDG Slicer Slicing Criteria Concern Graph PDG Constructor Jimple 3-Address Code Call Graph Points-to Set Soot Framework ( http://www.sable.mcgill.ca/soot/ ) SPARK Points-to Set Analysis Control-Flow Data-flow Analysis Annotated Jimple Translator

19. Department of Computer Science, Graduate School of Information Science & Technology, Osaka University Our effort to implement the system The program size  PDG Construction: 2731 LOC (without comments)  Slicing: 9296 LOC (without comments) slicing algorithms, heuristic functions and concern graph translation We could implement the PDG construction phase in two weeks:  One week to understand how Soot works.  The other week to implement code. Soot enabled us to focus on the essential part of the research idea.

20. Department of Computer Science, Graduate School of Information Science & Technology, Osaka University Advantage of Soot A rich analysis toolkit  Soot provides control-flow and data-flow for each method.  Jimple is simpler than source code and bytecode. Complex Java statements are simplified during compilation. Body Unit 1 n Stmt (Jimple code) is-a Value 1 n Expr is-a Control-flow Data-flow Local Exceptional UnitGraph SmartLocalDefs use Method

21. Department of Computer Science, Graduate School of Information Science & Technology, Osaka University Limitation of Soot  Soot is not a program analysis framework.  Soot keeps all data in memory to compile Jimple code to bytecode after the optimization.  Soot requires 2-4GB RAM to analyze jEdit and JDK.  Soot supports only the simple workflow: whole program analysis (call-graph construction) followed by local program analysis. We cannot implement a statistics tool (whole-program analysis) that uses the result of method-local analysis.

22. Department of Computer Science, Graduate School of Information Science & Technology, Osaka University Summary Concern location based on program slicing  We introduced heuristics in order to extract a functional concern of interest to a developer. Input is the same as a traditional program slicing.  Most of graphs can be printed on an A3-sized paper. Soot framework reduced the implementation effort.  Soot is a good framework, but we hope a framework specialized for program analysis. easy-to-learn, extensible and scalable