Speed comparisson between Kaffe/jdk on Linux/Windows

[Tim Wilkinson]

All,

Yeah the JIT is pretty slow - it's about 2 years old now and I haven't
tweeked it since doing the original version.

BUT!! Now the Beta 4 has been released I can say that Transvirtual hsa
been working on a new more optimized JIT which we'll be delivering into
the CVS tree as soon as we iron out a few problems (perhaps in a couple of
weeks).  This JIT uses the same final stage for generating code (so we
don't have to rewrite that) and the same initial stage (so we don't have
to rewrite that either) but the intermiediates are much more optimized -
this includes breaking out runtime checks to avoid filling the cache with
unuseful instructions, aliasing registers between stack slots to avoid
redundant copying, and inter-basic block register allocation ro avoid
spills/reloads in inner-loops.  Provisionally we're seeing a 4x speedup.

We still have inlining to do (not hard with the new framework) expecially
of JNI calls and <init> methods as well as provision for stack slot notes
to avoid such things as array bounds checks when the array size and
indexes are known (which happens quite a lot).

Will let you all know when this code is in the CVS so people can play.

In the meantime te GC still sucks (though thanks to Godmar and Jaon not as
much as it did) - we need a new strategy.

Cheers
Tim
Transvirtual Technologies, Inc.