The following benchmarks were run with the second version of the patch (so2.diff), once before and once after.

OSX 10.8 on a Macbook Pro, 2.5 GHz Core i5, two cores each with 256kb L2 cache, 3MB L3 cache:

1.32% less time for first benchmark (7.07 before, 6.98 after)
1.31% less time for second benchmark (9.81 before, 9.68 after)

Ubuntu 12.10 64-bit, AMD Athlon II X2 250 Processor (3.0 GHz, 2MB L2 cache, 256KB total L1 per processor):

5% less time for first benchmark (13.6 before, 12.9 after)
11% less time for second benchmark (18.3 before, 16.3 after)

Win7 64-bit, Intel Xeon E3-1230 V2 (quad-core, 3.30GHz, 256kb L2 cache per core, 8MB L3 cache):

1.8% less time for first benchmark (3.90 before, 3.83 after)
3.5% less time for second benchmark (7.27 before, 7.02 after)

However, this processor has "Intel SmartCache", which as far as I can tell means one processor can use other processors' caches if they are inactive. So I ran the benchmarks again with 4x larger data sets (revised benchmark script attached):

3.8% less time for first benchmark (19.3 before, 18.6 after)
7.5% less time for second benchmark (33.1 before, 30.6 after)