We have determined through testing that the vast majority of StructureIDs have only 0 or 1 transition. We should not create a whole HashMap in these cases.
Created attachment 24382 [details] Proposed patch It would be nice if someone else (Maciej?) ran benchmarks on his own machine to verify.
Comment on attachment 24382 [details] Proposed patch r=me But it would be nice if we could abstract away the single-transition optimization a bit, by having some helper functions that hide the core logic (e.g. one to get an existing transition for a property - it could check either the single entry or the hashtable as appropriate) etc.
Landed in r37627.