[Mono-list] JVM performance: JVM as a basis for CLR

Miguel de Icaza miguel@ximian.com
22 Jul 2001 20:40:00 -0400


Rhys,

> The reason why Miguel is seeing the Mono beat
> the pants off the JVM is very simple: VM size.

I do not. 

The reason I thought the CLR would be faster than the Java VM is
because back in my day I do not remember reading any limitations on
the use of the stack in Java.  Which do exist, but I ignored were
there.

I was wrong, as I said a few times already.

I saw the limitations on the use of the stack on the CLR as the right
thing to use a traditional instruction selector.  Having seen the
source code for various JITs and being unimpressed by the way they do
instruction selection, I figured that we could get better performance
than traditional JITs by implementing our JIT as a traditional
compiler.

I do not even care about the benchmarks that Paolo posted, because as
you point out, they are meaningless.  By the time things grow, by the
time you initialize all the static crap you link against, and the
gazillions of interactions you might have to deal with (how you
implement GC in a threaded environment) the story might be completely
different.

I think we can do better than traditional macro expanding JITs still
by using traditional instruction selection in our JIT.  

Miguel.