While I was investigating tsbs_query regressions with the new TCMalloc, I noticed that we can save one allocation for LimitSkipStage by having it create a ByteCode instance as a member variable.

Detailed information can be found here.

Other ideas like creating a per-plan ByteCode instance or per-thread instance were proposed but they are wider changes and will require broader perf impact analysis. On the other hand, this is a localized simple change and perf impact has been already analyzed (Links can be found at the above doc).