【正文】 are not executed in the same server or at the same time. Cond. 1 1 The total execution result of a master JSP page and the JSP fragments is the same as the JSP page before it was split. Cond. 1 2 Each JSP page, which is either a master JSP page or a JSP fragment, may be executed without other JSP pages. These conditions are satisfied by maintaining the data and control dependences existing in a JSP page after the JSP page has been split into fragments. However, it is hard work for developers to detect data and control dependent portions in a JSP page, because the data and control flow of a JSP page can be divided among the JSP page, the JSP engine that executes the JSP page, any Java beans used in the JSP page, and etc.. Thus, a natural solution calls for automatically detecting datadependent and controldependent portions linked to the portions that developers want to split. The program slicing technique detects a set of program statements that affect a program statement S by analyzing the data and control dependences existing in the program. One program slice consists of the detected set of program statements and the statement S. This technique is useful when debugging or maintaining large programs. Since JSP pages are piled into servlet programs by using a JSP piler such as the Apache Jasper piler [4], JSP slices can be obtained by applying program slicing technique to a servlet program piled from a JSP page. However, the program slicing technique is not sufficient to split a JSP page into fragments, because the technique does not take into account that program slices are executed independently. Figure 1 shows a part of a sample JSP page and a sample servlet program piled from the JSP page. Figure 2 (a) shows the data dependence graph obtained from the servlet program. In a program, when a statement S1 assigns a value to a variable and a statement S2 following S1 uses that value, a true data dependence exists between S1 and S2. Such a data dependence is denoted as S1 ? S2. In Figure 2 (a), due to the data dependences, the program slice for the statement S10 consists of S1, S9, and S10. If the program slice is cut from the servlet program, the remaining program statements have errors because the data dependences S1 ? S2, S11, and S12 and S9 ? S11 and S12 are not honored. Thus, the program slicing technique does not work well for splitting JSP pages. 3. JSP Splitting In this paper, we propose JSP splitting, a method of splitting a JSP page into fragments. The proposed method analyzes data and control dependences in the servlet program piled from a JSP page. At the time of analysis, JSP splitting takes into account that split JSP pages will be executed independently, and may not be executed in the same server or at the same time. First, the portions that developers want to split into a JSP fragment are selected. For the selected portions, the portions needed to satisfy the conditions 1 1 and 1 2 are detected. The detected portions are categorized into moved and copied portions, and are minimized in order to match the intentions of the developers. . Ignorable Data Dependences Basically, JSP splitting does not allow the dependent statements to be split apart. However, if this rule is applied too strictly, none of the statements shown in Figure 1 (b) can be split, because there are interlocking dependences starting from S1 and S9. Here, we consider how the variable out shown in Figure 1 (b) is used. The variable out is used to write the text data of any Web page. Note that the value of out can be retrieved from any JSP engine. When a master JSP page JM includes a JSP fragment JF by using a jsp:include tag and a JSP engine executes both of them, JM first retrieves the value of out from the JSP engine and writes its own text data to the value. Next, JM invokes JF. After being invoked, JF also retrieves the value of out and writes its own text data to the value. The text data written by JM is returned to out through the JSP engine and merged with the text data written by JM. This means that the value of out can be utilized by any JSP page and the text data written to the value is maintained by the JSP engines. In other words, when a JSP page is split into fragments, a statement that retrieves the value of out can exist both in a servlet program piled from a master JSP page and in servlet programs piled from JSP fragments. Therefore, the data dependences for the variable out can be ignored at the time of splitting a JSP page. As shown in Figure 2 (a), none of the statements can be split while maintaining the data dependences existing in the servlet program shown in Figure 1 (b). However, if the data dependences for the variable out are ignored, the data dependence graph is changed as shown in Figure 2 (b). Using the data dependence graph shown in Figure 2 (b) makes it possible to split S2 from the set of S9, S10, S11, and S12 which cannot be split due to the data dependences for the variable book. In servlet programs piled by the Apache Jasper piler, the data dependences for the variables shown in Table1 can be ignored for the same reason as for the variable the values of these variables are defined at the top of each servlet program and are frequently referred to, most statements cannot be split if the data dependences are to be maintained. In general, the outoforder execution is allowed for the statements without data and control dependences. Ignoring the data dependences for the variables shown in Table 1 may introduce an incorrect execution order of statements. For example,if S10 and S12 shown in Figure 1 are split into a JSP fragment and the JSP fragment is included before S11, an incorrect Web page is generated. In order to avoid