Curious Kaffe vs. jdk speed test results under Linux

[peter at transvirtual.com]

On Mon, 04 Jan 1999, Godmar Back wrote:
> My take on this is that the complete move the JNI only makes sense if
>accompanied by a move to a more advanced generational collector that
>actually requires JNI and compensates for it.

Let me take that one step further. I think we *have* to do it, mainly for
improving consistency and removing overhead (fighting entropy!), but it has to
be done in several directions:

(1) improve JNI call speed

(2) develop a new, *incremental* gc (the treadmill is my big favorite, now
that I think of a possible solution for the variable alloc size problem)

(3) re-structure the native method calls with respect to performance (passing
in builtin types instead of references, using return values, shadow frequently
accessed values in native land, .. like it is described in the "design" section
of <http://www.transvirtual.com/users/peter/native/talk.html>)

The various AWT libs gave me some experience with this transition (with the pure
kaffeh-based biss-kernel, the hybrid biss-only, and the pure JNI kaffe AWT).

-- Peter