http://gist.github.com/581704

Two major changes.

1. The fast_pack function now lives up to its name. Instead of constantly cat'ing data and reallocating buffers, the buffer is allocated once and data is safely copied in. A string is created once at the very end.

2. All uses of RARRAY_PTR were replaced with calls to rb_ary_entry(). This C API function is present on 1.8.7, 1.9.x and Rubinius. Using this instead of RARRAY_PTR prevents a lot of cache copying and flushing for Rubinius.

Performance under MRI was slightly improved when tested with bin/standard_benchmark. Improvement was approximately 5% for serializing/deserializing LARGE objects.