That seems to work. !5 seconds instead of 53 for the alternation is much more reasonable. Still doesn't beat the two pass method, but there is probably a break point somewhere.

c:\test>junk31 836952.log
Benchmark:
timing 1 iterations of
 AAtwo_pass, BBtwo_string, CCone_string
...

two pass: 10
AAtwo_pass:  7 wallclock secs ( 5.60 usr +  0.83 sys =  6.43 CPU) @  0
+.16/s (n=1)

            (warning: too few iterations for a reliable count)

two string: 10
BBtwo_string: 15 wallclock secs (14.35 usr +  0.94 sys = 15.29 CPU) @ 
+ 0.07/s (n=1)

            (warning: too few iterations for a reliable count)

one string: 7
CCone_string: 16 wallclock secs (14.76 usr +  0.73 sys = 15.49 CPU) @ 
+ 0.06/s (n=1)

            (warning: too few iterations for a reliable count)
[download]

I'm still puzzled why the two pass takes less time than the one string, but the numbers are right and I can't anything major wrong with the benchmark?