> Antoine suggests to relax the test and only check
> if the count is within a range (12, 13).

This feels like implementation driven testing.  I would prefer to disable the offending test pending further investigation into the core cause. 

It is not clear to me that these differences between 2.x and 3.x are not bugs.  It seems to me that the 2.x/3.x difference discussed in msg116336 above is simply manifestation of --with-computed-gotos becoming default in 3.x.