gcjx status

Mon Dec 5 16:49:00 GMT 2005

Mark asked me to send a note to the list with an overview of current
gcjx status.
* Front End
 We can parse and analyze 1.4 code just fine. There are some bugs
 (run against jacks to see) but so far they don't seem to be hit much
 by real code.
 There is one known parser bug for 1.5 -- with gcjx you can't
 intermix annotations and other modifiers. (This is just a mistake
 on my part.)
 A lot of semantic analysis for 1.5 works. With a few ugly hacks I
 can build the Classpath generics branch. Many error checks still
 need to be written. The simpler 1.5 features work ok, the problems
 are mostly with generics.
 We could benefit greatly from comprehensive jacks tests for all the
 new features. Some small number of tests do exist. At some point
 I will make a systematic effort to write more.
* Bytecode generator
 Aside from being slow gcjx is suitable for use as a 1.4 bytecode
 compiler today. I don't know of any bugs in the 1.4 part of
 bytecode generation. I've built Classpath and Mauve with it and run
 these with JamVM. I've also built Classpath with the
 verify-after-bytecode-creation flag, with good results. (So, it
 could benefit from heavier testing.)
 Much of the code needed for 1.5-style bytecode generation has been
 written. We still don't generate bridge methods, but I have been
 looking at that recently. The code to write enums, annotations,
 etc, exists but is largely untested.
* Tree back end.
 We can compile simple programs to object code and run them. My test
 cases have been the little programs in testsuite/libjava.lang. I've
 done this by building against a known-good libgcj.so, I haven't yet
 gotten a purely-gcjx-built libgcj to work. Compiling from .class
 does not work. I haven't done any work on 1.5 stuff here. I forget
 whether I've tested BC ABI compilation or not.
* Merge requirements.
 To merge I think gcjx needs to do what the current gcj does at least
 as well. Most of the needed work is in the tree back end, but also
 the parsing performance problem must be addressed, and the
 replacement for jcf-dump must be merged in. After a couple more
 generics-related tweaks I plan to focus on the tree back end.
Tom