[abc] revision of CC submission

Having contemplated the reviews of the CC paper, I think we want to
expand it by dumping the current sections 6 and 7, and instead have
a substantial section making a comparison to ajc which shows why
the architecture proposed here is a good thing. This will be needed
for SP&E and ECOOP alike....

6. Comparison with ajc
6.1 Source characteristics
    (use SLOCcount for this)
    - How big are ajc and abc?
    - What are the relative sizes of the type checkers?
    - What are the relative sizes of the weavers?
    - How well is ajc separated from the Eclipse compiler?
      How well is abc separated from Polyglot?
    - Why does ajc use its own version of BCEL?
      Is Soot also specific to abc (answer: no)

6.2 Compile times
    - benchmarks:
        3 pure Java (including abc itself)
        5 aspect (LoD-sim, LoD-weka, nullcheck-sim, profiler, ants)
        a few synthetic benchmarks, where we can artificially vary
         the number of pieces of advice and the size of the base program
    - measure:
        times for ajc and abc just running semantic checks
        times for ajc, abc -O0, abc

6.3 Runtime of object code
    - what is the effect of Soot optims?
        compare ajc, abc, ajc+Soot
    - brief summary of further optims, ref to PLDI submission

6.4 Ease of code generation and weaving
    - choice of rewriting at source versus IR:
      * thisJP -> thisJPStaticPart trafo in Polyglot
      (cf how done in ajc?)
    - advantages of Jimple for weaving
      * lots of locals in LoD for ajc
      * can rely on nullcheck elim when doing
        lazy inits
    - other examples?

6.5 extensibility
    - brief summary and ref to AOSD paper

7. Other related work

Implementation of AspectWerkz, JBoss, AspectC++.
Received on Sun Dec 12 14:26:32 2004