That said, it is possible to further improve json by reducing the number
of memory allocations and temporary copies. Here is an experimental
(meaning: not polished) patch which gains 40% in decoding speed in your
example (9 seconds versus 15).

We could also add an option to intern object keys when decoding (which
wins 400MB in your example); or, alternatively, have an internal "memo"
remembering already seen keys and avoiding duplicates.