Uploaded image for project: 'JDK'
  1. JDK
  2. JDK-8051619

JEP 201: Modular Source Code



    • Mark Reinhold
    • Feature
    • Open
    • Implementation
    • jigsaw dash dev at openjdk dot java dot net
    • L
    • L
    • 201



      Reorganize the JDK source code into modules, enhance the build system to compile modules, and enforce module boundaries at build time.


      This JEP does not change the structure of the JRE and JDK binary images, nor does it introduce a module system. That work is covered by the related JEPs 220 and 261.

      This JEP defines a new source-code layout for the JDK. This layout may be used outside of the JDK, but it is not a goal of this JEP to design a broadly-accepted universal modular source-code layout.


      Project Jigsaw aims to design and implement a standard module system for the Java SE Platform and to apply that system to the Platform itself, and to the JDK. Its primary goals are to make implementations of the Platform more easily scalable down to small devices, improve security and maintainability, enable improved application performance, and provide developers with better tools for programming in the large.

      The motivations to reorganize the source code include:

      1. Give JDK developers the opportunity to become familiar with the modular structure of the system;

      2. Preserve that structure going forward by enforcing module boundaries in the build, even prior to the introduction of a module system; and

      3. Enable development of Project Jigsaw to proceed without always having to "shuffle" the present non-modular source code into modular form.


      Current scheme

      Most of the JDK source code is today organized, roughly, in a scheme that dates back to 1997. In abbreviated form:



      • The share directory contains shared, cross-platform code;

      • The $OS directory contains operating-system-specific code, where $OS is one of solaris, windows, etc.;

      • The classes directory contains Java source files, and possibly resource files;

      • The native directory contains C or C++ source files; and

      • $PACKAGE is the relevant Java API package name, with periods replaced by slashes.

      To take a simple example, the source code for the java.lang.Object class in the jdk repository resides in two files, one in Java and the other in C:


      For a less trivial example, the source code for the package-private java.lang.ProcessImpl and ProcessEnvironment classes is operating-system-specific; for Unix-like systems it resides in three files:


      (Yes, the second-level directory is named solaris even though this code is relevant to all Unix derivatives; more on this below.)

      There are a handful of directories under src/{share,$OS} that don't match the current structure, including:

      Directory                     Content
      --------------------------    --------------------------
      src/{share,$OS}/back          JDWP back end
                      bin           Java launcher
                      instrument    Instrumentation support
                      javavm        Exported JVM include files
                      lib           Files for $JAVA_HOME/lib
                      transport     JDWP transports

      New scheme

      The modularization of the JDK presents a rare opportunity to completely restructure the source code in order to make it easier to maintain. We implement the following scheme in every repository in the JDK forest except for hotspot. In abbreviated form:



      • $MODULE is a module name (e.g., java.base);

      • The share directory contains shared, cross-platform code, as before;

      • The $OS directory contains operating-system-specific code, as before, where $OS is one of unix, windows, etc.;

      • The classes directory contains Java source files and resource files organized into a directory tree reflecting their API $PACKAGE hierarchy, as before;

      • The native directory contains C or C++ source files, as before but organized differently:

        • The include directory contains C or C++ header files intended to be exported for external use (e.g., jni.h);

        • C or C++ source files are placed in a $LIBRARY directory, whose name is that of the shared library or DLL into which the compiled code will be linked (e.g., libjava or libawt); and, finally,

      • The conf directory contains configuration files meant to be edited by end users (e.g., net.properties).

      • The legal directory contains legal notices.

      To recast the previous examples, the source code for the java.lang.Object class is laid out as follows:


      The source code for the package-private java.lang.ProcessImpl and ProcessEnvironment classes is laid out this way:


      (We took the opportunity here, finally, to rename the solaris directory to unix.)

      The content of the directories currently under src/{share,$OS} that don't match the current structure is now in appropriate modules:

      Directory                     Module
      --------------------------    --------------------------
      src/{share,$OS}/back          jdk.jdwp.agent
                      bin           java.base
                      instrument    java.instrument
                      javavm        java.base
                      lib           $MODULE/{share,$OS}/conf
                      transport     jdk.jdwp.agent

      Files in the current lib directory that are not intended to be edited by end users are now resource files.

      Build-system changes

      The build system now compiles one module at a time rather than one repository at a time, and it compiles modules according to a reverse topological sort of the module graph. Modules that do not depend on each other, directly or indirectly, are compiled concurrently when possible.

      A side benefit of compiling modules rather than repositories is that code in the corba, jaxp, and jaxws repositories can make use of new Java language features and APIs. This was previously forbidden, since those repositories were compiled before the jdk repository.

      The compiled classes in an intermediate (i.e., non-image) build are divided into modules. Where today we have:


      the revised build system produces:


      The structure of image builds, as noted, does not change; there are very minor differences in their content.

      Module boundaries are enforced at build time, insofar as possible, by the build system. If a module boundary is violated then the build will fail.


      There are numerous other possible source-layout schemes, including:

      1. Keep {share,$OS} at the top, with a modules directory to contain module class files:

      2. Put everything under the appropriate $MODULE directory, but keep {share,$OS} at the top:

      3. Push {share,$OS} down into the $MODULE directories, as in the present proposal, but remove the intermediate classes directory and prefix the names of the native and conf directories with an underscore, all so as to simplify the common case of pure Java modules:

      4. A variant of scheme 3, but with {share,$OS} at the top:

      5. Another variant of scheme 3, pushing {share,$OS} deeper down so as to further simplify the case of pure Java modules with no $OS-specific code:


      We rejected the schemes involving underscores (3–5) as too unfamiliar and difficult to navigate. We prefer the present proposal over schemes 1 and 2 because it entails the least change from the current scheme while placing all of the source code for a module under a single directory. Tools and scripts that depend upon the current scheme must be revised, but at least for Java source code the structure underneath each $MODULE directory is the same as before.

      Additional issues which we considered:

      • Should we define distinct directories for resource files, so that they would be separate from Java source files? — No; this does not seem worth the trouble.

      • Some modules have content that spans repositories; is this a problem? — It's an annoyance, but the build system can cope with it via the magic of the VPATH mechanism. Over time we might restructure the repositories to reduce or even eliminate cross-repo modules, but that's beyond the scope of this JEP.

      • Some modules have multiple native libraries; should we merge them so that each module has at most one native library? — No; in some cases we need the flexibility of multiple native libraries per module, e.g., for "headless" vs. "headful" AWT.


      As stated, this JEP does not change the structure of the JRE and JDK binary images, and makes only minor changes to the content. We therefore validated this change by comparing images built with it against images built without it, and running tests to validate the actual minor changes.

      Risks and Assumptions

      We assumed that Mercurial would be able to handle the massive number of file-rename operations that would be necessary to implement this change, and to preserve all historical information in the process. Early testing showed Mercurial to be capable of this, but there is still a minor risk that the relationships between the new and old locations of some files were not properly recorded. In that case the history of the file in its old location will still be in the repository; it will just be more difficult to find.

      It is impossible to apply a patch created against a repository using the old scheme directly to a repository using the new scheme, and vice versa. To mitigate this we developed a script to translate the file names in a patch from their old locations to their new locations.


      This JEP is the second of several JEPs for Project Jigsaw. It incorporates the definition of the modular structure of the JDK from JEP 200, but it does not explicitly depend upon that JEP.


        Issue Links



              alanb Alan Bateman
              mr Mark Reinhold
              Alan Bateman Alan Bateman
              Alan Bateman, Alex Buckley, Mandy Chung, Paul Sandoz
              Brian Goetz
              0 Vote for this issue
              16 Start watching this issue