slicing java 2023.pdf

Journal of Logical and Algebraic Methods in Programming 130 (2023) 100826 Contents lists available at ScienceDirect Journal of Logical and Algebraic Methods in Programming journal homepage: www.elsevier.com/locate/jlamp Program slicing of Java programs ✩ Carlos Galindo, Sergio Pérez, Josep Silva ∗ Valencian Research Institute for Artiﬁcial Intelligence (VRAIN), Universitat Politècnica de València, Camino de Vera s/n, E-46022 Valencia, Spain a r t i c l e i n f o Article history: Received 12 January 2022 Received in revised form 23 September 2022 Accepted 29 September 2022 Available online 3 October 2022 Keywords: Program slicing JSysDG Flow dependence Object-ﬂow dependence a b s t r a c t Program slicing is a technique to extract the part of the program that can affect the values computed at a given program point (known as the slicing criterion). To represent programs, program slicing uses the System Dependence Graph (SDG), for which several extensions like the Java System Dependence Graph (JSysDG) or the Sub-Statement Linear Dependence Graph (SSLDG) exist to deal with Java object-oriented programs. In this paper, we present an incompleteness result proving that these graphs do not produce complete slices in all cases, and speciﬁcally when some object variables are selected as the slicing criterion. We ﬁrst identify the source of the problem: the representation of dependences between partial deﬁnitions of objects is ill-deﬁned in these approaches, leading to a loss of completeness in many cases. To solve this limitation, we extend these representations with the addition of a speciﬁc ﬂow dependence for object type variables called object-ﬂow dependence. This extension provides a more accurate ﬂow representation between object variables and its data members and it allows us to obtain complete slices when an object variable is selected as the slicing criterion. © 2022 The Authors. Published by Elsevier Inc. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/). 1. Introduction Program representation is an important ﬁeld in the program analysis research area, and it is at the base of most program analysis and transformation techniques. An accurate representation of the internal program dependences is strongly associated with the quality, precision, and performance of many program analysis techniques. In this paper, we introduce a program representation for the analysis of object-oriented programs. Speciﬁcally, we propose a program slicing technique for object-oriented programs based on our new program representation. Program slicing [1,2] is a technique to extract from a program the set of statements, the program slice [3], that affects the value of a variable v at a given program point p ( p , v ), which is known as the slicing criterion [4]. Program slicing is applied in many disciplines such as software maintenance [5], debugging [6], and program specialization [7], among others. Example 1.1. Consider the Java code snippet in Fig. 1a, which contains a program with two classes: A and Main. The latter contains a main method that creates an instance of A. ✩ This work has been partially supported by grant PID2019-104735RB-C41 funded by Spanish MCI/AEI/10.13039/501100011033, by the Generalitat Valenciana under grant Prometeo/2019/098 (DeepTrust), and by TAILOR, a project funded by EU Horizon 2020 research and innovation programme under GA No 952215. Carlos Galindo was partially supported by the Spanish Ministerio de Universidades under grant FPU20/03861. Sergio Pérez was partially supported by Universitat Politècnica de València under FPI grant PAID-01-18. Corresponding author. E-mail addresses: [email protected] (C. Galindo), [email protected] (S. Pérez), [email protected] (J. Silva). * https://doi.org/10.1016/j.jlamp.2022.100826 2352-2208/© 2022 The Authors. Published by Elsevier Inc. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/). C. Galindo, S. Pérez and J. Silva 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 Journal of Logical and Algebraic Methods in Programming 130 (2023) 100826 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 class A{ public int x,y; public A (int a, int b) { x = a; y = b; } public void setX(int a) { x = a; } public int f(int a) { return a * y; } } class Main{ public static void main(String[] args){ A a1 = new A(1,2); int i = a1.f(10); a1.setX(3); A a2 = a1; } } class A{ public int x,y; public A (int a, int b) { x = a; y = b; } public void setX(int a) { x = a; } public int f(int a) { return a * y; } } class Main{ public static void main(String[] args){ A a1 = new A(1,2); int i = a1.f(10); a1.setX(3); A a2 = a1; } } (b) Slice w.r.t. 14,i (a) Java code Fig. 1. Java code fragment and slice with respect to variable i at line 14. If we observe that variable i at line 14 produces a wrong value, we can use program slicing to compute the part of the program that may have inﬂuenced this value, and thus may contain the bug. Fig. 1b shows the slice (the grey code is excluded from the slice) with respect to 14, i , represented with underlined code. Method f of class A internally uses its data member y to compute its result, but not data member x. Hence, as the slicing criterion is the result of a call to method f, the slice contains the deﬁnition of method f and tracks down where the value of y comes from. In this example, the value of y comes from the call to the A’s constructor in line 13. Although object a1 is included in the slice, we can exclude its data member x and its initialization, since its value does not affect the value of the slicing criterion 14,i. 1.1. The problem Currently, one of the most advanced program representations used to slice object-oriented programs is the Java System Dependence Graph (JSysDG) [8], a graph that can represent packages, interfaces, classes, and methods; and that gives support to polymorphism, dynamic binding, and inheritance. The slice shown in Fig. 1b is the minimal slice of the code in Fig. 1a, and it has been computed with the JSysDG. Although, in general, the JSysDG provides accurate slices, in this paper, we present an incompleteness result: we show that some scenarios exist where the JSysDG’s slices are not complete. Therefore, we do not reveal an imprecision problem (the slices contain more code than they need), but an incompleteness problem (the slices contain less code than they need), which means that some code that can affect the slicing criterion is not included in the slice. This is a fundamental problem because completeness is a property required by most applications of program slicing. For instance, the slice computed from a bug symptom could not contain the bug. Or the slice computed from a variable to produce a specialized program could produce a specialized code that is not equivalent to the original program. The source of the problem is that JSysDG is not prepared to select an object variable as the slicing criterion, and it can produce a slice where only some of its required data members are included. This lack of completeness is caused by the deﬁnition of ﬂow dependence in the JSysDG. The classic ﬂow dependence [1] was designed for variables that are atomically deﬁned or used in a single statement, but has never been reconsidered to deal with object variables, which can be partially deﬁned or used (e.g., by deﬁning or using only one of its data members) in a statement. Example 1.2 (JSysDG completeness counterexample). Consider again the code in Fig. 1a. We can deﬁne the slicing criterion 15, a1, which means that we are interested in the whole object a1 after1 executing the method call in line 15. This means that the slice should include all the data members of a1, and the code needed to deﬁne them. The code in black in Fig. 2a is the result obtained by slicing the JSysDG. The JSysDG includes the deﬁnition of a1.x in the slice because the function call a1.setX(3) is also included as part of the slice, but it ignores all the data members not being deﬁned there. As a result, data member a1.y in line 2 and its deﬁnition in line 5 are not included in the slice. Even the call to A’s constructor is missing. The result is an incomplete slice for 15, a1, which provides no value for data member a1.y, nor for a1 itself. The expected complete slice is the code in black of Fig. 2b. The problem described for the JSysDG is inherited by its later extensions, such as the Sub-Statement Level Dependence Graph (SSLDG) [9]. The SSLDG extends the theoretical model proposed by the JSysDG to accept more concrete slicing criteria. 1 In the rest of the paper, for simplicity in our examples, we interpret the slicing criterion as “a1 after executing line 15”. The complementary interpretation “a1 before executing line 15” can be achieved by simply changing the line of the slicing criterion (i.e., 14, a1). 2 C. Galindo, S. Pérez and J. Silva 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 Journal of Logical and Algebraic Methods in Programming 130 (2023) 100826 class A{ public int x,y; public A (int a, int b) { x = a; y = b; } public void setX(int a) { x = a; } public int f(int a) { return a * y; } } class Main{ public static void main(String[] args){ A a1 = new A(1,2); int i = a1.f(10); a1.setX(3); A a2 = a1; } } 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 (a) JSysDG slice class A{ public int x,y; public A (int a, int b) { x = a; y = b; } public void setX(int a) { x = a; } public int f(int a) { return a * y; } } class Main{ public static void main(String[] args){ A a1 = new A(1,2); int i = a1.f(10); a1.setX(3); A a2 = a1; } } (b) Expected slice Fig. 2. JSysDG and expected slices of the code in Fig. 1 w.r.t. 15, a1. Both models share the same representation for object variables in method calls, which is the key factor of the incompleteness problem. This paper presents an approach that solves the problem described in Example 1.2. We augment the JSysDG by replacing the current deﬁnition of ﬂow dependence with three more accurate deﬁnitions: the standard deﬁnition of ﬂow dependence for primitive variables and another pair of new deﬁnitions for object variables that we call object-ﬂow dependence and objectreference dependence. These deﬁnitions require the specialization of the ‘deﬁned variables (DEF)’ and ‘used variables (USE)’ sets of each statement, two concepts coming from the JSysDG that are explained later through Deﬁnitions 2.1 and 2.2. The specialization of these sets allows the JSysDG to distinguish when object variables are being totally or partially used or deﬁned in method calls. The rest of the paper is structured as follows. Section 2 recalls some key concepts about the construction of the JSysDG. Section 3 explains and justiﬁes the incompleteness problem of the current JSysDG when slicing an object that is the caller of a method call, illustrating it with an example in Java. Then, in Section 4, the DEF and USE sets for every statement are redeﬁned in order to properly represent deﬁnitions and uses of object variables and its data members. After that, Section 5 formally introduces object-ﬂow dependence and object-reference dependence, and justiﬁes their necessity in OO programs. Section 6 presents some slicing restrictions that need to be added to the slicing algorithm when traversing object-ﬂow dependences; and Section 7 presents the empirical evaluation comparing the JSysDG and the JSysDG extended with object dependences. Finally, Section 8 presents the related work and Section 9 concludes. 2. Background: the Java System Dependence Graph (JSysDG) This section has been included to keep the paper self-contained. Those readers already familiar with the JSysDG can skip this section, where we explain how the program representation ﬁeld has evolved to reach the current solution to properly represent the program in Fig. 1a so that we can automatically obtain the slice shown in Fig. 1b. In particular, we explain the JSysDG through its incremental evolution: CFG → PDG → SDG → ClDG → JSysDG CFG. The starting graph to build a JSysDG is the Control Flow Graph (CFG) [10]. It is a graph that represents all possible execution paths of a method. In the CFG, each statement is represented with a node, and two nodes are connected if they may be executed sequentially. Two nodes, Enter and Exit, are added as the initial and ﬁnal nodes of the method execution respectively. Additionally, every CFG node is augmented with two sets: the deﬁnition set and the use set, that denote the set of variables respectively deﬁned and used at this CFG node. Formally, Deﬁnition 2.1 (Deﬁnition set (DEF)). Let G be a CFG. Let n be a node in G representing a statement s. The deﬁnition set of n is denoted with DEF (n) and it contains all the program variables that are deﬁned (their value is assigned) at statement s. Deﬁnition 2.2 (Use set (USE)). Let G be a CFG. Let n be a node in G representing a statement s. The use set of n is denoted with USE(n) and it contains all the program variables that are used (their value is accessed) at statement s. PDG. From the CFG we can calculate two different dependences that are used to construct a Program Dependence Graph (PDG) [11]. These dependences are the control dependence and the ﬂow dependence, deﬁned hereunder. 3 C. Galindo, S. Pérez and J. Silva 1 2 3 4 5 6 7 8 9 10 11 Journal of Logical and Algebraic Methods in Programming 130 (2023) 100826 class A{ public int x, y; public A (int a, int b) { x = a; y = b; } public int getX() { return x; } public int getY() { return y; } public void f() { x = x + 1; } } class B extends A{ public B (int a, int b){ super(a,b); } public void f() { x = x + 2; } } 12 13 14 15 16 17 18 19 20 21 22 23 24 25 class Main{ public static void main(String[] args){ A a; if (Math.random() > 0.5) a = new A(1); else a = new B(2); a.f(); g(a); } public void g(A a){ System.out.println(a.getX() + a.getY()); } } Fig. 3. Fragment of Java code with a polymorphic call. Deﬁnition 2.3 (Control dependence). Let G be a CFG. Let n and m be nodes in G. A node m post-dominates a node n in G if every directed path from n to the Exit node passes through m. Node m is control dependent on node n if and only if m post-dominates one but not all of n’s CFG successors. Deﬁnition 2.4 (Flow dependence). Let G be a CFG. Let n and m be nodes in G. Node m is ﬂow dependent on node n if: (i) v ∈ DEF (n), (ii) v ∈ USE(m), and (iii) there exists a control-ﬂow path from n to m where v is not redeﬁned. The PDG of a method is a graph G = ( N , A ) where N is the set of nodes of the CFG minus the Exit node, and A is a set of arcs that represent control and ﬂow dependences. SDG. A program usually contains a set of methods connected by method calls. For this reason, in order to connect the PDGs of all the methods of a program and simulate parameter passing between calls and deﬁnitions, Horwitz et al. deﬁned the System Dependence Graph (SDG) [12]. An SDG represents each parameter of a method with a formal-in node, and a formalout node represents a parameter that may be modiﬁed inside the method. Analogously, each method call is augmented with an actual-in node for each argument of the call, and an actual-out node for each argument that may be modiﬁed by the method. The SDG connects method calls with their deﬁnitions representing parameter passing with (interprocedural) parameter arcs (input arcs connect actual-in with formal-in nodes and output arcs connect formal-out with actual-out nodes). Additionally, a call arc is generated to connect the call node to the method Enter node. Finally, a new kind of arc called summary arc is added to the SDG to describe the relation between input and output arguments in method calls. A summary arc connects an actual-in node and an actual-out node if the value related to the actual-in node is needed to calculate the value deﬁned in the actual-out node. ClDG. With an SDG as its base, the Class Dependence Graph (ClDG) [13] augments its representation to consider OO programs. The ClDG deﬁnes a class entry node for each class, connected to the method Enter nodes of all its methods by class membership arcs, and to all its data members by data membership arcs. In the ClDG graph, inheritance is represented with a class dependence arc from the base class to the derived classes. JSysDG. The Java System Dependence Graph (JSysDG) augments the ClDG with a representation for polymorphic calls and dynamic binding. This can be seen in Figs. 3, 4, and 5, which shows a Java program and two portions of its JSysDG that illustrate different aspects of the graph described above. This ﬁgure also shows how the JSysDG represents two speciﬁc scenarios that are worth mentioning to later understand the source of incompleteness: 1. A polymorphic object is the caller of a method, and the call’s target is only known at runtime. The JSysDG represents the caller and its deﬁned and used data members as a tree. There is a node for each possible dynamic type connected to the corresponding method deﬁnition. An example of this situation can be seen in Example 2.5. Example 2.5. Consider the code in Fig. 3, where a Java program with two classes with an inheritance relationship (B extends A) is represented. The program also contains a Main class with a main method that creates an object of either A or B dynamic type depending on a randomly generated number. In line 19, the method that would be executed in the call a.f() can only be determined at runtime, thus, the static representation of the program needs to deﬁne both possibilities. An example of how the JSysDG represents this method call is shown in Fig. 4. 2. A method call contains a polymorphic object as a parameter. In this scenario the JSysDG representation follows the proposal introduced by Liang and Harrold in [14], where the object parameter is represented in the graph as a tree 4 C. Galindo, S. Pérez and J. Silva Journal of Logical and Algebraic Methods in Programming 130 (2023) 100826 Fig. 4. JSysDG of call a.f() in line 19 of Fig. 3. Fig. 5. JSysDG of call g(a) in line 20 of the code in Fig. 3. structure, with a subtree for each possible dynamic class, unfolding all its data members in both method call and deﬁnition. This scenario is shown in Example 2.6. Example 2.6. Consider the call to method g in line 20 of Fig. 3. This call receives a polymorphic object (variable a) as a parameter. The corresponding JSysDG representation is shown in Fig. 5. In the JSysDG, only data members inside each type tree are linked, in order to accurately select only those data member used inside the method deﬁnition of g. Note that, when we follow the proposed representation, if an object is represented recursively, the tree representation may be inﬁnite. To address this issue, the JSysDG employs a k-limiting approach (a tree representation is only unfolded to a level k). 3. Limitations of the JSysDG As it has been proven with Example 2, even though being the current most accurate representation of Java OO programs, the JSysDG can produce incomplete slices. In this section we explain what its cause is. In Java, program variables can be of two different types. On the one hand, we have primitive variables, which are atomic (e.g., int i = 42). These variables are always deﬁned and used atomically, i.e., every time a primitive variable is deﬁned in the program, the new value of the variable replaces the previous one, and the previous value cannot be further accessed. 5 C. Galindo, S. Pérez and J. Silva Journal of Logical and Algebraic Methods in Programming 130 (2023) 100826 Fig. 6. JSysDG representing lines 13 and 15 of the main method in Fig. 1, and slice w.r.t. 15, a1 (grey nodes represent the slice, bold nodes the slicing criterion). On the other hand, there are object variables, which are compositionally formed by a collection of data members. Each data member, in turn, can be a primitive variable, or another object variable. The JSysDG was proposed by Walkinshaw et al. [8] extending the representation provided by Liang and Harrold for C++ [14]. Both approaches share the same deﬁnition for program slice and slicing criterion: “A program slice is a set of program statements and predicates that might affect the value of a program variable v that is deﬁned or used at a program point p; p , v is known as the slicing criterion.” According to this deﬁnition, an object variable can be considered as the slicing criterion. If we select an object variable as the slicing criterion, the slice should include all the statements that might affect the value of any of its data members. Nevertheless, as we have seen in Fig. 4, when an object variable is the caller of a method call, only the used and deﬁned data members of the object have a representation in the unfolding tree as argument-in and argument-out nodes respectively. Additionally, all the ﬂow dependences between object deﬁnitions and uses in method calls are propagated through their data members, but they never connect the node that represents the object variable itself. If we put all this knowledge together, we can identify a speciﬁc scenario where the slice obtained by the JSysDG is not complete when selecting some object variables as the slicing criterion. This scenario is given in the main method of Fig. 1, in particular in lines 13 and 15 (the relevant code is shown in Fig. 6). Line 13 creates object variable a1 of type A instantiating its data members x and y with a call to A’s constructor. Then, line 15 modiﬁes only the value of its data member x. Therefore, after line 15, the values of the data members (x and y) of object a1 come from two statements: x is deﬁned in line 15 (with value 3) and y is deﬁned in line 13 (with value 2). However, there is not ﬂow dependence between these two statements because method call a1.setX(3) does not depend on any data member of object variable a1 to deﬁne data member x. The described situation is represented in Fig. 6. Fig. 6 represents with grey nodes the slice produced by the JSysDG with respect to 15, a1. This slicing criterion is marked in the JSysDG with a bold line. Note that, when the slicing criterion is an object variable, then all the data members are marked as the slicing criterion as well. Thus, the slicing criterion is formed from the control dependence subtree of node a1. This slice corresponds to the code shown in Fig. 2a. Contrarily to the expected slice, which is the one in Fig. 2b, the slicing traversal of the JSysDG ends without reaching the declaration of a1, the constructor call new A(1, 2), and data member y in class A. This counterexample reveals that ﬂow dependences of object variables need to be redeﬁned so that slices are correctly computed when a whole object variable is selected as the slicing criterion. 6 C. Galindo, S. Pérez and J. Silva Journal of Logical and Algebraic Methods in Programming 130 (2023) 100826 Table 1 The ordered sequence DEF os and the USE set of some of the statements in the main method of Fig. 1a. Scenario Statement DEF os USE 1 3 2 A a1 = new A(1,2); a1.setX(3); A a2 = a1; [a1.x, a1.y, a1] [a1.x, a1] [a2.x, a2.y, a2] ∅ {a1} {a1} 4. Deﬁnitions and uses of object variables In order to solve the problem identiﬁed in the previous section, we need to extend the standard notion of ﬂow dependence (see Deﬁnition 2.4) for the case of objects. We start by giving a more accurate description for the DEF and USE sets when object variables are contained inside them. Unlike primitive variables, object variables are not completely replaced every time they are deﬁned. Since they are formed from a collection of data members, a statement may modify only some of them. As a result, the deﬁnition of all the data members of an object variable may be split into different statements. Hence, object variables can be deﬁned in two different ways: Deﬁnition 4.1 (Total deﬁnition of an object variable). An object variable v that points to a memory location m1 is totally deﬁned in a program statement s if the execution of s makes v to point to m2, and m1 = m2 (v points to a different object). Total deﬁnitions appear in two different scenarios: 1. An assignment of an object variable with a constructor call. This happens, e.g., in A a1 = new A(1,2) (see line 13 of Fig. 1), where a1 is totally deﬁned. Constructor methods always deﬁne2 all the data members of the variable and never use them. Hence, the DEF set for this statement (DEF (#13)) includes the object variable itself and all its data members. The order in which all these deﬁnitions occur is crucial for the deﬁnition of ﬂow dependence. For this reason, we transform the DEF set into an ordered sequence called DEF os where all data members are deﬁned ﬁrst, and the object variable is deﬁned at the end, e.g., the ordered sequence of deﬁnitions in the mentioned statement is DEF os (#13) = [a1.x, a1. y , a1]. In this scenario, USE(#13) = ∅ since no variable is used. 2. An alias assignment between two object variables that point to different objects. This happens, e.g., in A a2 = a1 (see line 16 of Fig. 1), where a2 is totally deﬁned.3 As in the previous scenario, the DEF set is transformed into an ordered sequence where data members are deﬁned ﬁrst, and the object variable is deﬁned at the end. In contrast, USE(#16) remains as a set of uses, and contains the object variable of the right-hand side of the assignment ({a1}). Deﬁnition 4.2 (Partial deﬁnition of an object variable). An object variable v is partially deﬁned in a program statement s if v points to the same object o before and after the execution of s, and s deﬁnes at least one data member of o. Partial deﬁnitions occur when a statement modiﬁes at least one data member of an object variable, but not the object it points to. Partial deﬁnitions appear in two different scenarios: 3. A method call that deﬁnes a data member of the caller. This happens, e.g., in a1.setX(3) (see line 15 of Fig. 1). Method calls may partially deﬁne object variables by deﬁning some or all of its the data members. The DEF set is also transformed to an ordered sequence which, as it happened in total deﬁnitions, includes all the data members deﬁned inside the method followed by the object variable itself. On the other hand, the USE set includes all the data members used inside the method together with the caller object itself, whose reference is needed to make the call. 4. An assignment of an object variable’s data member. This happens, e.g., in a1.x = 42. The JSysDG treats this case as a particular case of the previous scenario because the JSysDG inherits the program assumptions given by Liang and Harrold when building their version of the ClDG: “We further assume that data members of an object can be accessed only through a method (we replace a direct reference to a data member of an object as a method call)”. For this reason, before building the JSysDG, a1.x = 42 is a1.setX(42). Therefore, the DEF set turns into an ordered sequence that contains data member a1.x followed by the object variable itself. In this case, the USE set contains the object variable deﬁned (a1). Example 4.3. Table 1 provides an example of DEF and USE sets for the main method of the program in Fig. 1. In this table there are examples of the described scenarios. Deﬁnitions and uses appear in the order described in the previous scenarios. 2 3 This deﬁnition can be explicit or implicit, because Java initializes all data members by default. This type of total deﬁnitions include also assignments of the form A a = f(); where an object variable is deﬁned through the return of a method call. 7 C. Galindo, S. Pérez and J. Silva Journal of Logical and Algebraic Methods in Programming 130 (2023) 100826 Fig. 7. Input and output scopes in a method call. 1 2 3 4 (...) int x = 5; x = 1; y = x; // def x // def x // use x 1 2 3 4 5 (...) A a = new A(); a.setX(1); a.setY(2); b = a; // // // // def def def use a a a a (b) Def-use with object variables. (a) Def-use with primitive variables. Fig. 8. Flow dependence in presence of primitive and object variables. This new method to annotate deﬁnitions and uses of object variables in program statements is of fundamental importance for the redeﬁnition of ﬂow dependence. Moreover, it comes with an important advantage over the previous formulation: It allows us to differentiate the speciﬁc moment in which a data member and an object is deﬁned or used and, thus, it allows us to slice a program with respect to a speciﬁc instant in the computation (for instance, we can slice a.setX(3) with respect to the value of a ‘before’ or ‘after’ the method invocation). To incorporate this feature to the JSysDG, we have enhanced the graph representation for object variables at the scope of a method call. On the one hand, we provide an object tree for the scope to represent its value before the call (scope_in). The root of this tree represents the use of the object variable, and the tree contains the data members used inside the function call. On the other hand, if the object is modiﬁed during the method invocation, we provide another object tree that represents its value after the call (scope_out). The root of this second tree represents the deﬁnition of the object variable and it contains the data members deﬁned inside the function call. For instance, considering the function call a.setX(3), the resulting JSysDG is shown in Fig. 7. 5. Deﬁnition of object-ﬂow and object-reference dependences In Java, as it happens in most programming languages, a variable cannot be used without being previously deﬁned. The deﬁnition of a primitive variable is always total, but, as shown in Section 4, the deﬁnition of an object variable can be partial. This poses new diﬃculties in the computation of the usual ﬂow dependence. In fact, the standard deﬁnition (see Deﬁnition 2.4) is insuﬃcient. Let us illustrate the problem with a simple example. Example 5.1. Consider the code in Fig. 8a, where the primitive variable x is deﬁned twice and then used. According to Deﬁnition 2.4, x in line 4 depends on x in line 3, but x in line 4 does not depend on x in line 2. If we consider, however, the code in Fig. 8b, we have an analogous situation: a is deﬁned twice and then used. Therefore, according to the standard deﬁnition of ﬂow dependence, a in line 5 ﬂow depends on a in line 4 (which is correct), but a in line 5 does not depend on a in line 3 (which is clearly wrong). The rationale why Deﬁnition 2.4 does not work with object variables is because they can be partially deﬁned, while primitive variables are always totally deﬁned. This is also the problem of the JSysDG: it uses the standard deﬁnition of ﬂow dependence (Deﬁnition 2.4) with object variables. The solution is to extend the deﬁnition of ﬂow dependence to account for objects. Interestingly, this extension allows for novel situations. For instance, in Fig. 8b, a in line 5 depends on a in line 3 (even though it is redeﬁned in line 4), but also, a (partial) deﬁnition can depend on another deﬁnition. In particular, a in line 4, which is a (partial) deﬁnition, depends on a in line 3, which is another deﬁnition. With all these ideas, we identify (and formally deﬁne) two new dependences that complement the classical ﬂow dependence with a speciﬁc treatment for object variables (it does not apply to primitive variables, which continue using the classic 8 C. Galindo, S. Pérez and J. Silva Journal of Logical and Algebraic Methods in Programming 130 (2023) 100826 Fig. 9. CFG and eCFG nodes corresponding to statement A a1 = new A(1,2). ﬂow dependence deﬁnition (Deﬁnition 2.4)). We call these new dependences object-ﬂow dependence and object-reference dependence. The former can be formally deﬁned with an extended CFG. Deﬁnition 5.2 (Extended control-ﬂow graph). Given a CFG G = ( N , A ), its extended version eCFG is a graph G with the same nodes and arcs as G with one exception: every node n ∈ N that deﬁnes an object variable and contains m > 1 variable deﬁnitions (DEF os (n) = [ v 1 , v 2 . . . v m ]) is replaced by a sequence of nodes (n1 , ..., nm , connected by arcs) where each variable deﬁnition is represented by one single node. The eCFG is useful to explicitly represent the order in which a sequence of operations happen inside a single statement. These operations cannot be distinguished in a CFG because all of them are represented with one single node, while they can be distinguished in the eCFG because they are represented by a sequence of nodes. Example 5.3. Consider the assignment A a1 = new A(1,2) in line 13 of Fig. 1. This assignment contains a total deﬁnition of variable a1 and its CFG node together with its sequence of deﬁnitions and uses are shown in Fig. 9a. Note that the sequence of deﬁnitions appears in the order described by scenario 1 in Section 4. Since the node contains three different deﬁnitions, the node must be split into different nodes in the corresponding eCFG. Fig. 9b shows how the initial CFG node is split into three different nodes (one per variable deﬁnition) which are sequentially connected to form the corresponding eCFG representation. We can now formally deﬁne object-ﬂow dependence. Deﬁnition 5.4 (Object-ﬂow dependence). Let n and m be nodes in an eCFG. m is object-ﬂow dependent on n if: #1 (i) (ii) (iii) or #2 (i) (ii) (iii) n deﬁnes an object o, m uses the object o, and there exists a control-ﬂow path from n to m where object o is not redeﬁned. n deﬁnes a data member x of an object o, m deﬁnes object o,4 and there exists a control-ﬂow path from n to m where data member x of o is not redeﬁned. The ﬁrst set of conditions corresponds to the classic deﬁnition of ﬂow dependence, the deﬁnition-use dependence, which also applies to object variables, even if the deﬁnition is partial. The second one considers a deﬁnition-deﬁnition dependence, produced by partial deﬁnitions. This ﬂow dependence considers the case when the slicing criterion is an object variable, and it depends on all the complementary partial deﬁnitions that, together, produce the complete value of that variable. Object-ﬂow dependence represents all situations in which the values of the data members of an object are propagated. In particular, it is able to identify multiple partial deﬁnitions of an object and properly connect all of them to produce the whole value of an object. But object-ﬂow dependence does not consider the object reference of objects. Unlike primitive variables, object variables have a pointer to a memory position where a speciﬁc object is stored. This pointer is updated 4 I.e., in the eCFG, DEF os (m) = [ v ], where v is an object variable that points to object o. 9 C. Galindo, S. Pérez and J. Silva Journal of Logical and Algebraic Methods in Programming 130 (2023) 100826 Fig. 10. Code with data member redeﬁnition (left) and its JSysDG with object-ﬂow (right). every time a total deﬁnition of an object variable is executed (see Deﬁnition 4.1). Example 5.5 shows a scenario where object-ﬂow dependence is insuﬃcient to include the reference of an object in a slice. Example 5.5. Consider the code in Fig. 10, where an object a of class A (which has one single data member x) is totally deﬁned in line 2. Then, line 3 redeﬁnes data member x (even though its unique data member is redeﬁned, the object reference remains unchanged). Finally, line 4 contains a use of variable a. All object-ﬂow dependences in this code are represented in the graph of the ﬁgure labeled with #1 and #2. These object-ﬂow dependences are the ones generated by the ﬁrst (#1) and second (#2) sets of conditions in Deﬁnition 5.4, respectively. With these dependences, the slice computed for 4, a is shown with grey nodes. Clearly, this slice is incomplete, because line 2 should be included in the slice. The problem is that there is a missing dependence between a_out in line 3 and a in line 2. The missing part is the object reference of a, which is deﬁned in line 2. The rationale is the following: even though all data members of a_out are deﬁned in line 3, the object reference of a_out is not. Therefore, a_out in line 3 must depend on the code that deﬁnes its object reference (line 2). To account for these missing dependences, we deﬁne a new type of dependence called object-reference dependence that complements object-ﬂow dependence. It connects objects with their object reference. Formally, Deﬁnition 5.6 (Object-reference dependence). Let G be a CFG. Let m and n be nodes in G. n is object-reference dependent on m if m totally deﬁnes an object o, n partially deﬁnes object o, and there exists a control-ﬂow path from m to n where object o is not totally deﬁned. Object-ﬂow and object-reference dependences are the key elements to solve the incompleteness problem of the JSysDG. The completeness of both dependences with respect to any deﬁnition-use combination of statements with object variables is an important result that we formulate here and prove in Appendix A and Appendix B. Theorem 5.7 (Object-ﬂow completeness). Let s1 ; s2 ; . . . sn ; be a sequence of statements. Let x be an object variable or a data member of an object variable deﬁned at s1 . If the value of an object o at sn data depends on the execution of s1 , then there is a transitive object-ﬂow dependence between s1 and sn . Theorem 5.8 (Object-reference completeness). Let s1 ; s2 ; . . . sn ; be a sequence of statements. Let x be an object variable totally deﬁned at s1 . If the reference of an object variable v at sn depends on the total deﬁnition of s1 , then there is a path formed by object-ﬂow and/or object-reference arcs between s1 and sn . The ﬁnal JSysDG that includes our new dependences is shown in the following example. With object-reference dependence it solves the problem explained in Example 5.5. Example 5.9. Consider again the code in Fig. 1a. Fig. 11 shows its associated JSysDG. The upper part of the ﬁgure corresponds to the representation of class Main, where we can see how object dependences are also deﬁned over the tree structure of method calls. Although object dependences add arcs between object variables, the original deﬁnition of ﬂow dependence is still applied to primitive variables. This happens in the call a1.f(10), where data member y of the constructor call is 10 C. Galindo, S. Pérez and J. Silva Journal of Logical and Algebraic Methods in Programming 130 (2023) 100826 Fig. 11. JSysDG of the program in Fig. 1 and slice w.r.t. 15, a1. linked to the argument-in node y_in. We can see that the object-reference arc is needed to connect the object variable a1 after the call a1.setX(3) with the construction of this object in call new A(1,2). This JSysDG extended with object dependences can properly slice the graph from any slicing criterion. For instance, the slice computed for the slicing criterion 15, a1 in Fig. 1a would be the expected slice, i.e., the code in black in Fig. 2b. 6. The slicing algorithm The JSysDG slicing algorithm is the standard one proposed by Horwitz et al. [12]. It computes slices in two phases: (i) traverse the graph backwards from the slicing criterion collecting all nodes reached using any arc except for output arcs, (ii) traverse the graph backwards from any node in the slice collecting all nodes reached using any arc except call and input arcs. The overall process has a linear time complexity. This algorithm can be used with our graph, producing the same precision as with the JSysDG. However, this algorithm does not take advantage of the new object-ﬂow dependences, producing a loss of precision. The standard algorithm would include in the slice all the data members of an object variable even if we are only interested in one of them. For instance, in Fig. 11, if we consider the node y inside the method call ne w A (1, 2), the #2 algorithm would unnecessarily include in the slice data member x and its value 1 (due to arc a1 ←− x). The problem can be solved by limiting the traversal of the object-ﬂow arcs in certain cases. When the traversal reaches a node n, an incoming object-ﬂow arc can only be traversed if one of the following three conditions is true: 1. n has been reached via an object-ﬂow arc, 2. n is the slicing criterion, or 3. n is a predicate. Condition 1 ensures that the traversal of object-ﬂow arcs is still transitive. Conditions 2 and 3 provide a starting point to traverse object-ﬂow arcs. When the slicing criterion is an object variable (Condition 2), we need to traverse all object-ﬂow arcs to include in the slice the value of all its data members. On the other hand, in Condition 3, when the traversal reaches a predicate (e.g., the condition of an if or a while statement) we need to follow object-ﬂow arcs to include the data members used by its condition in order to keep the slice complete. The restriction imposed to the traversal of object-ﬂow arcs increases the precision of the algorithm in the presence of object variables, while keeping its linear time complexity. Algorithm 1 includes all these conditions in the original slicing algorithm proposed in [12], making it suitable to traverse the JSysDG after adding object-ﬂow dependences. 11 C. Galindo, S. Pérez and J. Silva Journal of Logical and Algebraic Methods in Programming 130 (2023) 100826 Algorithm 1 Slicing algorithm for the JSysDG with object-ﬂow dependence. Input: A JSysDG G and the slicing criterion node nsc . Output: The set of nodes that compose the slice S of G w.r.t. nsc . Initialization: S 0 ← {nsc , none}. 1: function MarkNodesOfSlice(G , nsc ) 2: S 1 ← MarkReachingNodes(G , S 0 , nsc , {Output}) 3: S 2 ← MarkReachingNodes(G , S 1 , nsc , {Call, Input}) 4: S ← {n | n, arcType ∈ S 2 } 5: return S Phase 1 Phase 2 6: function MarkReachingNodes(G , N , nsc , ArcTypes) 7: WorkList ← N 8: while WorkList = ∅ do 9: select some n, lastArcType ∈ WorkList 10: WorkList ← WorkList \ n, lastArcType 11: for all arc ∈ getIncomingArcs(n) do 12: if arcType ∈ ArcTypes then 13: continue 14: m ← getSourceNode(arc) 15: if arcType = ObjectFlow then 16: if lastArcType = ObjectFlow ∧ n = nsc ∧ ¬isPredicate(m) then 17: continue 18: N ← N ∪ m, arcType 19: WorkList ← WorkList ∪ m, arcType 20: Change ( A) Change (B) return N Algorithm 1 illustrates the slicing process by including a couple of relevant changes to Horwitz’s algorithm. The ﬁrst one ( A) is that the slicing criterion nsc is now used as a parameter of function MarkReachingNodes, since this function needs to identify in Line 16 whether the current node is the slicing criterion (to check Condition 2). Additionally, (B) the WorkList in function MarkReachingNodes contains now tuples of two elements (see Line 9) instead of a single element, storing also the information about the type of the last traversed arc (to check Condition 1). The three aforementioned conditions are checked in lines 15-17. If any of them holds then the current node is not included in the slice. Example 6.1 shows how the application of Algorithm 1 solves the incompleteness problem of the JSysDG presented in Section 3. Example 6.1. The graph in Fig. 11 represents the JSysDG of the code in Fig. 2 augmented with object-ﬂow and objectreference dependences. It shows the slice computed with respect to the object variable a1 in line 15 after the method call a.setX() (the slicing criterion node is marked with a bold line in the graph). To clearly show the two traversal phases of the algorithm, the slice is divided into two sets of nodes. The nodes marked in light grey are added to the slice during Phase 1. On the other hand, the dark grey nodes are the nodes added to the slice during Phase 2. The slice code is exactly the expected slice shown in Fig. 2b. It is worth mentioning that the change proposed in Section 6 to Horwitz et al.’s algorithm has a direct impact on the slice’s precision. Note that the traversal algorithm improves its accuracy by preventing the x data member of object variable a1 in method call new A(1,2) to be included in the slice. This happens because x can only be reached from a1 (through an object-ﬂow arc) and this arc can only be traversed from another object-ﬂow arc, according to Condition 1 of the slicing algorithm. Even though a1 can be reached from two object-ﬂow arcs, none of them belong to the slice, thus x is never included in the slice. 7. Empirical evaluation In this section we compare our implementation, which include object-ﬂow and object-reference dependences, with the original JSysDG. We have compared both graph implementations by measuring the graph generation time, the slicing time, and the size of the slice, comparing the results obtained by both slicer executions. All the algorithms and ideas described in this paper have been implemented in a prototype slicer for Java programs called JavaSlicer. JavaSlicer is open source and publicly available.5 To compare the performance difference between the JSysDG and JavaSlicer, we used the publicly available Java library re2j, a library developed by Google to work with regular expressions in Java. In particular, we used the most recent release of this library (version 1.66 ). re2j is a project with 8100 lines of code, distributed in 19 different Java ﬁles. In order to evaluate the techniques proposed throughout this work, we used both the JSysDG and JavaSlicer to build and slice the whole re2j library. Then, to make the comparison meaningful, we selected as slicing criteria all the object variables (root node of the 5 6 Available at https://github.com/mistupv/JavaSlicer. Available at https://github.com/google/re2j/releases/tag/re2j-1.6. 12 C. Galindo, S. Pérez and J. Silva Journal of Logical and Algebraic Methods in Programming 130 (2023) 100826 Table 2 Summary of experimental results, comparing the slices of re2j produced by the JSysDG (A) and JavaSlicer (B). Size range SCs Slice time (A) Slice time (B) Slower Size (A) Size (B) Improv. [0, 100) [100, 1000) [1000, 1400) [1400, 1800) [1800, ∞) [0, ∞) 49 95 122 146 31 443 0.076 ± 0.001 ms 130.98 ± 1.823 ms 622.67 ± 10.833 ms 881.09 ± 13.101 ms 1603.05 ± 11.769 ms 602.13 ± 9.078 ms 0.927 ± 0.018 ms 217.315 ± 2.874 ms 1164.423 ± 13.093 ms 1584.023 ± 12.429 ms 2943.965 ± 15.702 ms 1095.440 ± 12.039 ms 48.876 0.782 0.857 0.779 0.842 6.126 9.10 608.69 1284.66 1535.62 1994.42 1130.99 39.59 738.68 1664.99 1948.16 2522.74 1439.91 693.74% 23.19% 29.32% 26.77% 26.73% 100.47% Build times: JSysDG: 10.85 ± 0.09 s, JavaSlicer: 13.35 ± 0.07 s (23% slower). object tree representations) that are directly connected to a return statement.7 The result of this selection is a total of 443 different slicing criteria located over the 19 Java ﬁles. All experiments were done on an Intel Core i5-7600 processor with 16 GB RAM (DDR4 at 2400MT/s) under a Linux OS with kernel version 5.18.1. The experiment was run with the OpenJDK Java Virtual Machine (version 11.0.15+10). During the execution of the experiment, all processes and services except for the program slicer and the shell it was launched from were completely stopped to prevent interferences in the CPU performance. The methodology used to measure the performance of both slicers was the following: each time measurement (the graph generation or a speciﬁc slice) was repeated 101 times, and the ﬁrst iteration was always discarded (to avoid inﬂuence of dynamically loading libraries to physical memory, data persisting in the disk cache, etc.). Finally, we computed the average with error margins (with 99% conﬁdence) with the remaining 100 values. The results of the experiments performed are summarized in Table 2.8 In it, the slicing criteria are grouped according to the size of the JSysDG slice (in nodes). This is due to the strong inﬂuence the size of the slice has on the time required to compute it and on the relative sizes of slices produced by both graphs. The columns of Table 2 are described as follows: • • • • • • Size Range: the sizes of the JSysDG slices (in nodes) that are contained in each row. SCs: the number of slicing criteria contained in a particular range. Slice Time (A/B): the average time required to slice the corresponding JSysDG (A) and JavaSlicer (B). Slower: how much slower is JavaSlicer in comparison to the JSysDG. It is computed as (Time B − Time A )/Time A . Size (A/B): the average number of nodes in the JSysDG (A) and JavaSlicer (B) slices. Improv. (improvement): the average increase in size of the JavaSlicer slice compared to the JSysDG slice. It is computed as (Size B − Size A )/Size A . To interpret the results, we ﬁrst need to focus on the usage of the SDG. Typically, a graph will be built once and sliced multiple times, and thus its creation can be (and often is) much more costly than its traversal. Regarding complexity, creation is polynomial and traversal is linear. Our results are as expected: creating the graph is between one and four orders of magnitude more time-consuming (depending on the ﬁnal slice’s size). Regarding the graph creation, we can observe a 23% increase between the JSysDG and JavaSlicer, which is attributable to the great number of object trees that are featured on the graph. The graph itself consists of 42000 nodes, most of which represent objects, their polymorphism or their members. Thus, computing object-ﬂow and object-reference arcs represents a noticeable increase in time consumption. If we turn our attention to the slicing portion of the results, we can see that the slices can be classiﬁed into two distinct groups. The ﬁrst group consists of slices whose size is below 100 in the JSysDG tend to “blow up”, as the addition of object-ﬂow and object-reference adds a signiﬁcant number of nodes to the slice (relative to the size of the original slice). Thus, the relative columns of the ﬁrst row of results show a slicer that includes almost 700% more nodes and takes about 50 times longer. Due to the small size of the resulting slices and the short time it takes the JSysDG to compute them (0.08 ms), a 50x increase only manages to bring the average up to 0.9 ms, which is negligible in most applications of program slicing. This ﬁrst group can be considered an outlier, and it affects the averages of the table as a whole. The second group is represented by the rest of the table (slices with sizes above 100 nodes), where the results show that slices increase between 23% and 30%, with the cost being a 80% time increase. Throughout the table, time has increased linearly with slice size, as can be seen in Fig. 12. Whether the cost is worth the improvement in completeness is up to the application of slicing. For example, in compiler optimization, completeness is an important requirement for slicers; while for dependency highlighting or light debugging, a faster but incomplete slice may be more appropriate. Another important remark is the comparison between the slice results computed over the two models. We veriﬁed that every JavaSlicer’s slice included all the statements of the associated JSysDG slice. The rationale behind this is that the graph representation used by JavaSlicer is a conservative extension of the JSysDG (i.e., it also contains all the nodes and arcs of the JSysDG), where object-ﬂow and 7 Note that, to compare the performance of both models, selecting an object variable during the slicing traversal is preferable because any slicing criterion that produces a slice in the JSysDG without object variables would produce the same slice in our augmented JSysDG, as the two representations are identical except for object-ﬂow and object-reference arcs. 8 The full dataset produced by the slicer execution is publicly available at https://mist.dsic.upv.es/git/program-slicing/SDG/-/snippets/9. 13 C. Galindo, S. Pérez and J. Silva Journal of Logical and Algebraic Methods in Programming 130 (2023) 100826 Fig. 12. Relationship between the size and time required to compute a slice (logarithmic scale). object-reference arcs are added. Therefore, for any possible slicing criterion, the slice computed with the JSysDG (S JSys ) will always be contained in the slice computed with JavaSlicer (S JS ): S J S ys ⊆ S JS 8. Related work In 1996, Larsen and Harrold [13] proposed the ﬁrst graph representation able to slice OO programs, the Class Dependence Graph (ClDG). Their approach was the ﬁrst to provide a way to represent inheritance, polymorphism, and dynamic binding. The ClDG connected all the methods and data members of a class in a single graph. Their proposal was later improved by other approaches like those of Tonella et al. [15], or Liang and Harrold [14]. Although all these proposals focused on OO programs, none of them noticed the difference between total and partial deﬁnitions in object variables and their impact in ﬂow/data dependence. When we review the literature looking for speciﬁc slicing approaches for Java, we ﬁnd interesting related papers. For instance, Hammer and Snelting [16] deﬁned a new object unfolding process for method calls in presence of recursive data types, improving the results of the k-limiting approach proposed in [14]. They deﬁned an algorithm to completely unfold object variables without unfolding the same object twice in the same object tree and without loosing dependences based on point-to information. Kashima et al. [17] compared four different backward slicing techniques for Java: static execute before (SBE), based in the CFG; context-insensitive slicing (CIS), ignoring the call context; hybrid model (HYB), where the slice is deﬁned as the intersection of SBE and CIS; and improved slicing (IMP), based on the Hammer and Snelting’s work [16]. They compared their precision, scalability, and tradeoffs, determining IMP as the more accurate but not applicable to large programs. Other techniques focused on the representation of polymorphic objects in Java. This is the case of the JSysDG [8] and the SSLDG [9], mentioned at the beginning of this paper. Both graphs include a multiple-layer representation that contemplates Java programs at different levels: package level, class level, method level, and statement level. At statement level, both graphs use the unfolding of object variables in method calls using a tree-like representation for data members proposed in [14]. There are three main differences between these two graphs: the ﬁrst one in that the SSLDG includes a sub-statement layer where object variables outside method calls are further split into their data members, making the representation of direct accesses to object ﬁelds explicit in the graph; the second one is that the SSLDG proposal enhances the information of the slicing criterion when slicing polymorphic objects, using a triplet as slicing criterion where the new element indicates the dynamic type of the object being sliced to further reduce the computed slice; and the third difference is the modiﬁcation of the slicing algorithm to adapt its graph for forward slicing. Other works dealt with object-oriented programs in other programming languages like C++ or Python. In [18], Pani et al. present an algorithm for ﬁnding dynamic slices for object oriented programs in presence of function overloading on C++; and in [19], Jain and Poonia proposed a mixed static and dynamic slicing for C++ OO programs, where they generate dynamic slices in a faster and more accurate way by using object-oriented information in C++. These works are centered on dynamic slicing, and they are not focused in the representation of object variables, but in taking advantage of the information given by the compiler to compute the slice. On the other hand, a program slicing approach for Python is presented by Xu et al. in [20]. This work deﬁnes the Python System Dependence Graph (PySDG), used to slice Python First-Class objects. The PySDG takes all the ﬁrst-class objects including functions, methods, classes and modules into consideration, to construct the dependencies between the deﬁnition and use statements of these ﬁrst-class objects. In this model, the authors also introduce a new kind of dependence (Entity Dependence), which describes the dependency relationship between the statement which deﬁnes the entity object and the statement which calls the entity object. There are also some other works mainly focused on modeling data dependences for slicing in object-oriented programs. Chen and Xu [21] augmented the PDG of each method with tags, used to distinguish the different deﬁnitions and depen14 C. Galindo, S. Pérez and J. Silva Journal of Logical and Algebraic Methods in Programming 130 (2023) 100826 dences inside a statement. The authors deﬁned ﬁve different sets: Def (s), Ref (s), Def (s, x), Dep_D(s, x), and Dep_R(s, x). These sets were used to annotate data dependence arcs with the program variables involved in each data dependence. Despite the perspective is interesting, its purpose is different to ours. It gives extra information to data dependences by annotating them, limiting the graph traversal at slicing level when reaching a node looking for just a particular variable. Contrarily, our approach focuses on establishing a dependence between an object variable and the value of all its corresponding data members in any program point. The work by Orso et al. [22] exhaustively analyzes data dependence in the presence of pointers. They considered two different aspects: the classiﬁcation of deﬁnitions and uses, and the classiﬁcation of different kinds of paths in the CFG. In their work, they differentiate 24 kinds of data dependence and allowed the possibility of slicing with respect to only some of them including the considered dependences as part of the slicing criterion. Unfortunately, their data dependence is more suitable for point-to analysis than for the OO paradigm, as it is based on point-to relationships. Our approach may seem similar to object slicing, introduced by Liang and Harrold [14], or class slicing, deﬁned by Chen and Xu in [23], but there are some differences between their approaches and ours. In object slicing, the slicing criterion is deﬁned with a tuple v , p , o where v is a variable in a program statement p, and o is an object variable of the program. Object slicing determines which statements of o’s class affect the slicing criterion through object variable o. First, a standard slice is computed for the criterion v , p . Then, considering the computed slice, a new process isolates all method calls with o as the caller object. Afterwards, method deﬁnitions corresponding to o’s detected method calls are identiﬁed. Finally, o’s slice is computed by extracting from the initial slice all the statements corresponding to those method deﬁnitions. Note that, in object slicing, the slicing criterion is not the object variable itself, but a mechanism to reduce the statements included in the original slice. Our objective is different: we are interested in considering the object variable o as the slicing criterion, extracting from the whole progra

slicing java 2023.pdf

Document Details

Tags

Related

Full Transcript

Upgrade to continue