Arena::allocateFrom is slower than it could be. There are two reasons for this:

* Allocation requests are followed by memory zeroing, even though the client plans to overwrite the memory contents with the contents from some other data source.

* Allocation requests cause some code in Bits::reserveMemory to be executed. This code, which was designed to periodically "kick" the GC when using direct buffers, is not only slow, but also a bottleneck when working with multiple threads.