I found through profiling that batch_list_exact and batch_dict_exact spent a fair amount of time looking up the pickling function. I have found a way to optimize the lookup using a simple heuristic that assume items in dicts and lists have often a common type. So we have save some time by jumping right to the pickling function we need. Here are the result of running the benchmarks: ### fastpickle ### Min: 0.593904 -> 0.541918: 1.10x faster Avg: 0.606811 -> 0.564606: 1.07x faster Significant (t=10.18) Stddev: 0.01061 -> 0.02733: 2.5760x larger Timeline: http://tinyurl.com/bwbvw7j ### pickle_list ### Min: 0.907891 -> 0.915905: 1.01x slower Avg: 0.970537 -> 0.948349: 1.02x faster Significant (t=2.08) Stddev: 0.07107 -> 0.02526: 2.8136x smaller Timeline: http://tinyurl.com/cmlymq7

In addition, we could bring back the switch dispatch based on the first letter of the type name. It does seem to speed up things as well but as much as the type cache optimization.

I found a bug in the pickle_list benchmark. These are the new results: ### fastpickle ### Min: 0.603416 -> 0.542480: 1.11x faster Avg: 0.628639 -> 0.575296: 1.09x faster Significant (t=6.29) Stddev: 0.03330 -> 0.04985: 1.4969x larger Timeline: http://tinyurl.com/c7xm7p6 ### pickle_list ### Min: 0.366062 -> 0.317290: 1.15x faster Avg: 0.391959 -> 0.331500: 1.18x faster Significant (t=15.86) Stddev: 0.02131 -> 0.01651: 1.2910x smaller Timeline: http://tinyurl.com/cad2mkx

Why did you add all those calls to PyMemoTable_Get?

Are you asking why do we need to call both PyMemoTable_Get and memo_get? Or, why do we fetching the memo was moved to the save functions? For the former, there is no real reason. The extra call could be removed though profiling doesn't show this call as a hot spot. If the latter, we need this because the patch allow save(), which is where the memo was fetched, to be skipped.

Thanks for the explanation. I guess I don't like the code duplication that this patch adds. Perhaps a macro hiding the "memo_get" code blocks? Also, adding your explanation as a comment would be nice too.

The patch is too complicated for too little.