Andreas,

  First of all, I'm supposed to try BranchPrediction-ar.2.cs on a
stock VM, right?  I just don't see that much difference you reported
on my Pentium-M laptop.  Here is the result:

-------------------------
10	0	0
100	0	0
1000	1	0
10000	3	3
100000	30	26
1000000	295	258
10000000	2934	2569
100000000	29361	25694
-------------------------

  About 13-14% speed up.

  I just tried it on the XScale PXA 255 machine.  The results from
wellPredicted and misPredicted are almost the same.  I guess the
access to the main memory access hides the predictors.  (There are
only 36 bytecodes/branches in the inner loop, so I think the number of
branch target buffer entries on XScale is more than enough.)  Here is
the results from XScale 255 400MHz:

-------------------------
10	0	0
100	1	0
1000	3	2
10000	21	21
100000	214	219
1000000	2141	2141
-------------------------


-- Yoshiki