Method dispatch has shown a 76% regression on recent builds. 

This is on the bm_vm2_unif1_1.0 test:

i = 0
def m a, b
end

while i<6000000 # benchmark loop 2
  i+=1
  m 100, 200
end