Benchmark: timing 1000000 iterations of index, s///, substr...
     index: 15 wallclock secs (14.02 usr +  0.00 sys = 14.02 CPU) @ 71
+326.68/s (n=1000000)
      s///: 13 wallclock secs (12.33 usr +  0.00 sys = 12.33 CPU) @ 81
+103.00/s (n=1000000)
    substr: 14 wallclock secs (13.37 usr +  0.00 sys = 13.37 CPU) @ 74
+794.32/s (n=1000000)
          Rate  index substr   s///
index  71327/s     --    -5%   -12%
substr 74794/s     5%     --    -8%
s///   81103/s    14%     8%     --
[download]

So I guess it must be highly platform-dependent then. I'm using perl 5.8.0 on Mac OS X (darwin 6.3) on a PPC G3 (iBook).

Assuming you're using x86, a possible explanation would be that doing regex might benefit from the PPC's rich number of general-purpose registers, while the x86's memory-oriented instruction set is perhaps more suited at executing the (stack-based) perl opcodes.

But it's just a guess...

•Update: I'm enclosing all my benchmark stuff in readmore-tags to avoid cluttering the thread