Tableswitch performance

In r15045 I modified the implementation of tableswitch in the baseline and opt compiler for Intel. This resulted in a satisfying speed up to many benchmarks, in particular a near 15% speed up of luindex [1]. Currently switches of >= 8 values are converted into lookupswitches (ie explicit compare and branches for each case). On Intel having 2 branches within 16bytes is problematic for the branch predictor. We should reevaluate the choice of the constant 8 for lookupswitches in the light of having a better tableswitch implementation and that we might be polluting the branch predictor with too many branches within a small period.

[1] http://jikesrvm.anu.edu.au/cattrack/results/habanero.anu.edu.au/perf/5612/production/default/perf-dacapo/luindex-3/luindex-3-time.large.png