Eclipse Goes Native
Eclipse is an open-source, extensible integrated development environment (IDE) that's growing quickly in popularity. Written in Java, it provides a multilanguage development environment that allows developers to code in Java, C and C++. In response to the need for improved performance and additional platform coverage for the Red Hat Developer Suite, of which Eclipse is the core, we created a version of Eclipse that's compiled natively. Instead of running on top of a virtual machine the way Java programs usually do—although that can still be done if the user prefers—Red Hat's version of Eclipse is compiled to binary and runs natively using the libgcj runtime libraries, similar to the way a C program runs using the GNU C libraries.
To compile Eclipse natively, Red Hat's Eclipse Engineering team used GCJ, a free, optimizing, ahead-of-time compiler for Java. GCJ can compile Java source code to native machine code, Java source code to Java bytecode and Java bytecode to native machine code. The approach we took involves using GCJ to compile Java bytecode to native machine code.
This article discusses why native compilation was an attractive choice; explains what we had to do to GCJ, libgcj and Eclipse to make it possible; and shows, using a real-world example, that open-source Java has come a long way and now is useful commercially.
Two main factors from the early days of Developer Suite planning and engineering drove us toward native compilation: platform coverage and performance. Red Hat Enterprise Linux was scheduled to ship on several 64-bit architectures, and we wanted to make sure Developer Suite could run on all of them. One big problem was Eclipse had never been run on a 64-bit platform and it contained some code, specifically the interface between SWT, the graphics toolkit in Eclipse, and its native C libraries, that assumed 32-bit addresses. Aside from having to create a clean 64-bit version of SWT, we were faced with a more significant problem: no 64-bit Java Virtual Machine (JVM) for x86_64, AMD's 64-bit architecture, existed at the time, and it didn't look hopeful that one would be available before we had to ship.
Another problem we had was performance. Eclipse worked well on Microsoft Windows but the version available at the time was pretty slow on Linux. We found that startup alone took well over a minute, and early user testing found that the interface was a little too sluggish for comfortable use. For example, Eclipse is based on perspectives, which are collections of views and editors, only one of which is visible at a time. Switching between them is something that a user does fairly frequently. However, changing perspectives introduced substantial delays we thought unacceptable for the enterprise development market Red Hat Developer Suite was targeting.
The solution we came up with was to use GCJ to compile Eclipse into native binaries that could run without having a JVM installed. We knew that native compilation would help with the performance problems, because we would no longer have the overhead that comes with the JVM layer. It also would solve the platform coverage problem, as GCJ/libgcj was available on all of the 64-bit platforms we had to support, although in some cases, such as x86_64, it still needed a lot of work. Native compilation solved the technical problems we had and gave us the additional benefits of reducing our external dependencies, allowing us to make some significant improvements to open-source Java and to demonstrate that open-source Java has matured to the point of being useful commercially.
At the outset of this project, we really didn't know if it was possible to compile Eclipse with GCJ and expect it to run. First, Eclipse is a large program—more than two million lines of code as counted by wc. We didn't know much about Eclipse internals or what runtime facilities it might use. Second, GCJ's background is in embedded systems, and we knew that work remained on parts of the Java programming language, class loaders in particular, which are used heavily by Eclipse. Third, the free class libraries were not complete. We didn't know if Eclipse could use facilities we hadn't written yet or even whether Eclipse might break the rules and use internal, undocumented com.sun.* interfaces, as too many Java programs seem to do.
We therefore took a two-pronged approach to determining whether a project like this could succeed. First, we used GCJ to make a list of the APIs used by Eclipse that we did not or could not implement. To accomplish this, we wrote a shell script that would try to compile each Eclipse Java archive library (jar file) to object code. We then looked through the error messages to see what was missing. The results of this script were not encouraging: we found a large number of missing packages. Still, more investigation was required because some things didn't make sense. For instance, there were dependencies on the Swing graphical user interface classes, but we knew that Eclipse used SWT and not Swing.
Further investigation showed that many of the weird undefined references came not from Eclipse itself but from the third-party jar files included with it. For example, Eclipse includes its own copy of the Ant build tool and its own copy of the Apache Tomcat dynamic Web server. We knew that in many cases, the referenced classes would not actually be invoked in the Eclipse environment. This encouraged us to take another look at how to get Eclipse working.
Our second angle of attack was to try running Eclipse using the bytecode interpreter that comes with libgcj. By doing this, we reasoned, we would concentrate on runtime bugs, including the aforementioned class loader problems and missing functionality actually used by Eclipse.
This approach also was discouraging initially. We ran into problems not only with class loading, but also with the fact that libgcj's implementation of protection domains needed work. These are the bases for Java's secure sandbox architecture, which allows untrusted code to be run in a secure way. Problems in this area had an unfortunate shadowing effect—we had to fix each bug before we could discover the next one.
Our first round of changes to libgcj was bug fixing only. We implemented protection domains properly. Then, we made a pass over the entire runtime, fixing bugs related to class loading. Because of the way class loading had been implemented in libgcj, we had to modify all the places in the native code that conceivably might load a class to forward the request to the appropriate class loader.
Once this was done, we were able to start Eclipse using the libgcj bytecode interpreter. At this point the question became, how can we take real advantage of GCJ to compile Eclipse?
The naïve approach to this dilemma, namely precompiling all the classes and linking them all together, had been ruled out by our investigations into Eclipse's internals. This approach would clash with Eclipse's relatively sophisticated class loading strategy.
More investigation revealed that most classes are loaded by instances of the DelegatingURLClassLoader, which is a subclass of the standard URLClassLoader that has been extended to understand Eclipse's plugin architecture. It seemed like the best approach was to modify Eclipse to allow it to load precompiled shared libraries as well as bytecode files. We reasoned that the required changes would be localized due to the way plugin class loading had been structured.
In fact, we had to go one step further and extend libgcj a bit as well. libgcj knew how to load shared libraries invisibly in response to a call to, for example, Class.forName(). However, this magic always happened at the level of the bootstrap class loader. That wouldn't work well for Eclipse or for any other application that defines its own class loaders, so we invented a new gcjlib URL type. This is like a jar URL, but it points to a shared library. We also made some minor extensions to our implementation of URLClassLoader so that gcjlib URLs would be treated specially.
Doing this wasn't enough, however. We also had to solve the linkage problems. In particular, if we compiled a jar file to a shared library, how could we prevent the dlopen() of such a shared library from immediately failing due to unresolved symbols? The solution to this problem was to resurrect and clean up the -fno-assume-compiled option in GCJ. This option, which never had been finished, enabled an alternative ABI that caused GCJ's output to resolve most references at runtime rather than at link time.
The -f-no-assume-compiled option has various limitations and inefficiencies. On the boards for the future is a cleaner way to achieve this same goal. On the GCJ mailing list (see the on-line Resources section) this option is referred to either as the binary compatibility ABI or -findirect-dispatch. This new ABI does everything -fno-assume-compiled does, but in a much more efficient and compatible way. Development is underway and is coming along nicely on this new feature, one of several contributing to GCJ's enterprise readiness.
Once all this was in place, we finally were ready to make our changes to Eclipse. These turned out to be remarkably small. Most of the work involved making the same sort of change in three different places. In essence, we modified Eclipse so that when it's looking for a plugin's jar file, it also looks for a similarly named shared library installed alongside it. If there is one, we rewrite the URL passed to the class loader from a jar URL to a gcjlib URL. All rewriting is done conditionally, so our natively compiled Eclipse still works with an unmodified JVM. In other words, users are not locked in to native compilation if they would rather use a JVM instead.
Once that was done, we wrote our own launcher that understood how to bootstrap the Eclipse platform from shared libraries. This was accomplished in a modest 90 lines of code.
After all that, Eclipse was mysteriously slow. Had we done something wrong? Was GCJ-compiled code substantially worse than the code generated on the fly by the current crop of just-in-time (JIT) compilers? Did -fno-assume-compiled have enormous overhead?
One nice advantage of GCJ is its output generally can be treated in the same way one treats any object code. That is, existing tools such as OProfile can be applied to it directly without any change. And that, in fact, is how we investigated our performance problem.
The first thing we noticed was a large number of exceptions being thrown during platform startup. Amid the grumblings of compiler writers (exceptions should be for exceptional circumstances), and although we were considering changes to the GCJ runtime that would violate Java semantics, we noticed a strange symbol in the OProfile output. It turned out that a small bit of buggy assembly code deep in the libgcj runtime was causing a linear search of exception handling tables rather than the expected binary search. The overhead of this search through the entire program every time an exception was thrown was vast. A fix to the errant assembly code proved this was the problem, and suddenly our natively compiled Eclipse was able to start a second faster than the stock version using a JVM. To quantify it a bit further, the startup time dropped from more than a minute before the fix to less than 15 seconds after it.
Currently, we don't compile Eclipse directly from source to object code. Instead, we compile to bytecode and then compile the jar files to shared libraries. This is done for two reasons. First, a few bugs in the GCJ source compiler haven't been fixed. Second, Eclipse comes with its own build scripts that compile from source to bytecode. Reworking the Eclipse build system to allow building directly from source to binary seemed like a much larger divergence from the upstream sources than we were willing to maintain.
Also, we currently don't precompile all the jar files to shared libraries—some remain as jar files and are interpreted at runtime. This is done because the class libraries still are incomplete, and these jar files refer to classes that have not been implemented yet.
One of our patches is unsuitable for the public GCJ. We had to disable the compile-time bytecode verifier, as it was too buggy to compile some of the Eclipse jar files. We're in the process of replacing this verifier with a more robust one.
In addition, one limitation of natively compiled Eclipse deserves mention. You can't use natively compiled Eclipse to debug a GCJ-compiled application, because JDWP, the Java Debug Wire Protocol used by Eclipse, hasn't been implemented in libgcj yet.
The achievement of the native compilation of Eclipse is a strong indication that open-source Java based on GCJ and libgcj/classpath has reached the point of being commercially useful. That said, it's still not complete. Some fairly substantial gaps still need to be filled in before open-source Java can be a proper drop-in substitute for proprietary JVMs.
One of the major areas that needs work is the development/integration of a JIT compiler. JIT would allow a GCJ-based open-source Java environment to be used in a manner similar to a conventional JVM, meaning that native compilation and platform-specific binaries would not be necessary for performance reasons.
The other major piece that needs work also is, by far, the most visible missing piece—Swing. Work on an open-source implementation of Swing is coming along nicely as part of the GNU Classpath Project, but Swing is a huge undertaking and the GNU Classpath implementation is still not quite usable.
A full-featured and completely open-source Java environment is an attractive alternative to proprietary JVMs, and it's now within reach. During the past six months, Red Hat has more than doubled the number of engineers working in support of the Open Source Java solution and community. Eclipse is a large, complicated piece of software, and natively compiling and running it was an excellent test of and testament to the progress being made on open-source Java. The power of open source lies in its communities, so please consider joining the open-source Java community and contributing to the GCJ and GNU Classpath Projects in any way that interests you.
Resources for this article: /article/7549.
John Healy is the manager of Red Hat's Eclipse Engineering group, based in Toronto (people.redhat.com/jhealy). In the past he's worked on custom open-source toolchains for embedded processors as well as CRM and computer-telephony applications.
Andrew Haley has been a programmer for longer than he cares to remember. He is one of the maintainers of GCJ. He works for Red Hat, which supports him in this task.
Tom Tromey has worked on free software since the early 1990s. Patches of his appear in GCC, Emacs, GNOME, Autoconf, GDB and probably other packages he has forgotten about. He works at Red Hat as the technical lead of the Eclipse Engineering team. He can be reached at firstname.lastname@example.org.