Interesting. I guess I'd naively assumed that the 'bytecode engine'
would have some mechansim of "registering" certain classes to act as the
key classes it requires. (My hypothetical setup here is the tuned
hotspot engine but no java.* class files whatsoever)
What we've learnt however is that its going to be very hard to create a
lightweight platform built on the hotspot JVM. Anyone trying to do this
is going to need to ship a fair chunk of the Java SE class libraries.
Splitting the Java SE platform like that would also break the
JCP/platform rules, again making it less likely. Thus, it seems we will
always be shipping the entire Java SE platform even for a new language
that essentially wants to use none of it.
I guess that modular Java helps a little bit here in minimising the
required class file list?
Stephen
Jeremy Manson wrote:
Even with just a bytecode execution engine, you would need pretty much
all of the Throwables in java.lang - NullPointerException,
OutOfMemoryError, ArithmeticException, VerifyError, ClassFormatError,
IndexOutOfBoundsException and the like.
The flip side of "what classes does the VM need" is "what VM support
do the classes need". You can't actually implement a compliant
version of Java without special VM support for a bunch of the classes.
You would have to know what all of those requirements were, too.
Jeremy
On Fri, Aug 21, 2009 at 1:24 AM, David Holmes - Sun
Microsystems<david.hol...@sun.com> wrote:
Stephen,
Stephen Colebourne said the following on 08/21/09 18:14:
If you wanted to code from scratch a JVM, but not include the rest of the
Java SE platform (such as the .class files), what would you need to include?
Is Object.class mandatory for a pure JVM? Anything else?
The answers given imply that the JVM is not an isolated well defined
component. Yet that argues against being able to write a new programming
language for the JVM that has no integration whatsoever with the Java
language and just emits 'pure bytecode' for a 'pure JVM'.
I think you would need to separate the bytecode-engine part of the JVM from
the "Java runtime platform" part. The bytecode engine has few dependencies
on classes, perhaps not even Object, while the runtime environment has many
dependencies.
David Holmes
I, like Carsten, find it odd if this isn't well-defined.
Stephen
David Holmes - Sun Microsystems wrote:
As Martin stated what you are looking for is not part of the JVMS nor the
JLS, but the platform specification, which is essentially the entire set of
Java API's as found for example here:
http://java.sun.com/javase/6/docs/index.html
But the implementation of those classes will then have dependencies on
other implementation specific classes, so I'm not sure that you will be able
to establish a transitive closure of required classes that is independent of
the JVM in question.
If you were thinking about this from a basic language perspective - eg we
must have Object, and we must have Class, and array implies Serializable
etc, then there is a core set of classes that form the transitive closure of
the JVM bootstrap process. If you are interested in that then
-XX:+TraceClassLoading (might need a debug VM) will give you the set used by
a particular VM. But again this list is dependent on how those classes are
implemented themselves, so the list is JVM dependent.
HTH
David Holmes
Martin Buchholz said the following on 08/21/09 03:32:
The set of all public APIs that must be part of the java se platform
are tested by the platform tck, in particular by the "signature test",
and you can get the sources for that test (for research only)
and from that it should be possible (with work) to get a list of
all required classes. But that's a very large list, so is probably not
what you want. In practice, the subset of rt.jar of public classes
matching java.* or javax.* is a pretty good approximation.
Martin
On Thu, Aug 20, 2009 at 08:26, Carsten Otto
<o...@informatik.rwth-aachen.de <mailto:o...@informatik.rwth-aachen.de>>
wrote:
Hello,
I am working on automated termination analysis of Java Bytecode and I
am
missing an important bit of information in the Java Virtual Machine
Specification (JVMS). I'd be happy to get some help from you!
Every JVM needs to provide certain classes including code for their
native
methods, e.g. java.lang.Object (obvious) and java.io.Serializable
(because
every array implements it). The list of such classes can easily be
extended, but I have huge problems finding a lower bound to keep
this list
as small[1] as allowed according to the JVMS.
Is there some part of the specification that states which classes
need to
be provided? I can only see references to the API (e.g. "reflective
APIs"),
but no clear definition of the classes that need to exist. In the
current
draft for JVMS 3rd edition, the necessity to include
java.io.Serializable
is not even part of the JVMS, this is only visible by looking at the
definition of arrays in the Java Language Specification (JLS).
Some hints in this direction are also appreciated. So far I can only
guess,
that sun.awt.* is not part of the language defined in the JVMS - but
who
knows?
[1]: Sadly, just using whatever Sun or OpenJDK provide in rt.jar
does not
work.
Thanks a lot,
--
Carsten Otto o...@informatik.rwth-aachen.de
<mailto:o...@informatik.rwth-aachen.de>
LuFG Informatik 2 http://verify.rwth-aachen.de/otto/
RWTH Aachen phone: +49 241 80-21211
-----BEGIN PGP SIGNATURE-----
Version: GnuPG v1.4.9 (GNU/Linux)
iEYEARECAAYFAkqNaygACgkQjUF4jpCSQBTSEwCaA+MU0U73Cx1QS3Mgvr/6ZRET
4vcAnicKi99CFGZe1kE+jcUXD3bz9m4J
=Brg8
-----END PGP SIGNATURE-----
