Allocation Without Locking
In a programming environment with both concurrency and automatic garbage collection, the allocation and initialization of a new record is a sensitive matter: if it is interrupted halfway through, the allocating process may be in a state that the garbage collector can't understand. In particular, the collector won't know which words of the new record have been initialized and which are meaningless (and unsafe to traverse). For this reason, parallel implementations usually use a locking or semaphore mechanism to ensure that allocation is an atomic operation. The locking significantly adds to the cost of allocation. This paper shows how allocation can run extremely quickly even in a multi-thread environment: open-coded, without locking.