I don't think it's a 5.6 thing: I ran eduardo's code on ActiveState 5.6 with these results:

Benchmark: running not traditional, traditional, each for at least 10 
+CPU seconds...
not traditional: 11 wallclock secs (10.68 usr +  0.02 sys = 10.70 CPU)
+ @ 364259.16/s (n=3899030)
traditional: 11 wallclock secs (10.85 usr +  0.00 sys = 10.85 CPU) @ 4
+74001.38/s (n=5141493)
[download]

Judging by the difference, the compiler doesn't seem optimize that kind of switch.