RE: [ObjectWeb architecture] just an idea for (Java) native comp ilation

["Leskov, Dmitry" <dleskov@xxxxxxxxxxxxxxxxx>]

> >     There are a few other issues when native compilation can make a
> > difference, such as the download size of a self-contained package. Have
> a
> > look at http://www.excelsior-usa.com/jetbenchsize.html - download sizes
> > start from as little as 0.34 MB ("Hello, world" JetPerfected against JRE
> > 1.3.1 and zipped.) 
> 
> from your experience, does that factor scale linear with application size
> or is that difference in "Hello World" bound to some static offset? I
> mean, maybe with large applications that offset will not be the problem as
> its influence will shrink compared to the influence of the linear code
> size.
> 
        First of all, the size factor depends not exclusively on the
application's size but on which subest of the platform APIs it uses. The
Java 2 API classes reference each other in a spaghetti fashion, and the
situation becomes worse with each J2SE release. On JRE 1.4.2, even the
minimal "Hello, world" sample involves 696 (!) classes from rt.jar, of which
248 are actually loaded - just run "java -verbose hello". The global
optimizer of course removes unused methods, inserts stubs instead of classes
that are imported but never instantiated, and so on, otherwise the resulting
EXE would have been much larger.

        Then, the classes that constitute the Swing and AWT APIs are very
tightly coupled, and there are many of them. As a result, an executable for
even the smallest GUI app would have a size of several megabytes, whereas a
large non-GUI program concentrated on something like XML transformation may
compile to a smaller executable. The use of SWT also helps in that regard.

        We also have the internationalization issue. If you need your
compiled application to support all locales and encodings supported by the
respective JRE, you have to include all the locale classes and all classes
from charsets.jar, some of which, such as Japanese or /Chinese/Korean
contain very large tables 

        To summarize (and to answer your question), the factor scales linear
if we define "application size" as "the total size of application's classes
plus all required classes from the JRE". So the global optimization benefits
for a large application are likely to be limited to performance.