YCSB find/update regresses a bit after BSONElement construction was made more expensive. There are several opportunities to mitigate performance impact here:

• Make the iterators store raw pointers instead of BSONElements. Iteration does size computation instead of constructing a whole BSONElement.
• BSONElement constructor can save loads. Performance is extra sensitive here so we should look to merge loads where possible. There are 2 identified ways we can do this:
  • Merge the fixed size table and the variable size mask
  • Merge variable size branch with the computeRegex branch