The most instructive part of this are the listings from the changelogs alongwith the resulting improved performance. I'm really enjoying the C++ tunings. I wish they did show the results with the google hashes, though.
So, the C++ performance numbers were with the "public" version? It wasn't clear whether they left the internal version out completely (besides the changelog) or simple didn't release the code. The published C++ numbers are already way better than any other language in terms of speed and memory use. It would be surprising if there was a substantially faster version.
The final list of changes is supposed to be with the internal version, and roughly seem to show a 20% change over and above what the basic std::map -> ext/hash_map replacement. I'm not sure if I should add the percentages, or multiply them along to get the final runtime, though that doesn't make much difference, really :)
You're right that there may not be much improvement, but I've seen the google dense_hash_map do better sometimes than the basic SGI hash_map.
Edit: the code repo has the unoptimized versions. Oh well, at least I can get the satisfaction of seeing the 30% myself :)
Isaac, do you think you could add this problem to the shootout? I believe (and am therefore almost certainly wrong) that there isn't any benchmark there that derives almost purely from compiler theory.
