Hello folks,
three days ago I came across ATL and wanted just to try it out. Since I
could not get all necessary components (ATL, AM3 and the TCS editor)
from an update site, I resolved to check everything necessary out from
CVS and build it on my own.
It took me three nights to get the stuff running in a second Eclipse
instance and a lot of debugging the classloader action, so I thought I
should share my finds. I hope I did not forget anything in the following
list which is reconstructed from memory and by looking at the diff
between my workspace and the CVS:
- First I found org.eclipse.m2m.atl.dsls included an outdated version of
the org.eclipse.gmt.tcs.extractor/injector packages still expecting
TCSInjector while the code calling it was passing in a TCSRuntime
I checked out AMMACore and AMMAScript to be able to use the build.xml
provided in org.eclipse.m2m.atl.dsls to update the package to the
latest versions of extractor/injector and update the ASMs at the same
time.
- Thereafter I found org.eclipse.m2m.atl.engine had the same problem
(TCSInjector instead of TCSRuntime). I ended up updating the
respective jars there as well and had to update the ATL_Parser.jar
there as well (took the one from org.eclipse.m2m.atl.dsls.
- Now I had problems with the outline view. I needed to update
org.eclipse.m2m.atl.services.core/ServiceTransformationUtil.java from
branch 1.3.4.1 to resolve TCSInjector issues.
- Last I could not use Inject KM3 to Ecore because it was unable to
locate the org.eclipse.gmt.tcs.injector.KM3_ANTLR3* classes. This is
because the parser/lexer classes are in the same package which also
houses TCSRuntime, but TCSRuntime and the parser classes are exported
by different bundles (namely org.eclipse.m2m.atl.dsls and
org.eclipse.m2m.atl.engine).
Equinox cannot handle merging exports of the
same package from different bundles - no way. I tried buddy exports,
boot loading, whatever and finally found that there is no way and it
is good so because otherwise one may end up with a bad mix of
different versions of class files at runtime. I resolved to put a
copy of KM3-parser.jar into the org.eclipse.gmt.am3.dsls.km3 bundle.
Before I tried the above, I thought if might be a good idea not to have
the injector/exporter packages jarred up add all - why not simply depend
on the bundles. When I tried that, it introduced a cyclic dependency! It
seems that the bundles are not cut very well here. Similar goes for the
parser/lexer classes which I believe should be moved to their own
bundles so they can be used from multiple places. Furthermore I believe
the parse/lexer classes should be moved to a package
org.eclipse.gmt.tcs.injector.XXX where XXX would be ATL, KM3, etc. and
should be either moved to their own bundles thereafter to be usable by
different other bundles or may be included in various other bundles but
they must not be exported by those to avoid classloader confusion at
runtime.
Hope this was a helpful analysis.
I am not sure if this went to the right newsgroup. I did not want to
start cross-posting. If this is relevant to other newsgroups, please
somebody forward it or tell me to post it there.
Cheers,
Richard