Curious Kaffe vs. jdk speed test results under Linux

[Godmar Back]

>
> JIT, in general, should be faster than interpreted.  However, there is
> overhead: the JIT must compile the code to begin with.  If it takes longer
> for the JIT to run than you save on the code, then overall, JIT will be
> slower.  That's why newer VMs such as Sun's HotSpot only JIT when they "know"
> that a method will be frequently used.

Well, if we ever get Senthil Kumar's efforts on integrating intrp and jit,
kaffe will have that too.

>
> As far as the Windows JIT goes, another technique is that all of the standard
> class libraries can be pre-JITted and stored on disk.  Kaffe does not
> preprocess anything in java.*, so we JIT everything from scratch each time.
> (We should have an option to pre-JIT the standard class libraries.  I even
> have a wierd idea of how it might work.)
>

The not-so-weird way how this will work is by using egcs's gcj Java front-end.
Other groups (HP India) have already reported success in combining egcs-
generated code with their VMs.

With Cygnus's run-time library being rooted in an earlier version of Kaffe 
(Per Bothner used to be contributor to Kaffe), this should certainly be
possible.  Tim tried it and reported partial success a few months ago.

If anybody wanted to tackle this project, it would be very welcome.
At this point, the difficulty is that Cygnus hasn't published the interfaces
to their run-time lib yet, so you'll have to look hard at the egcs source and 
at the assembly code generated by egcs to find out what they are.
Also, exception handling is done differently.  But that would be a very,
very useful project.  For extra credit, create documentation of the
interfaces.

     - Godmar