Some applications benefit a lot by reusing existing cons cells instead of actual consing. Usually this technique is called hash consing. Users of this technique are e.g. the optimizer of the computer algebra system Maxima and the theorem prover ACL2.
This particular implementation is intended for use cases where performance is so critical, that even a single hash table access per cons is too expensive. To achieve such near-optimal speed, this library does not actually provide conses, but uconses. A ucons has not only a car and a cdr, but also a table of past users. Furthermore, the cdr of each ucons is restricted to other uconses or NIL. This setup has several advantages:
Checking whether a certain ucons already exists is a single lookup of its car in the table of its cdr.
The immutability of a ucons is enforced by its defstruct definition
The compiler has reliable type information of the slots of a ucons.
Lists of uconses are neither circular, nor improper.
Unfortunately there is also a painful downside of this approach. Traditional cons cells are a fundamental Lisp data type and well supported throughout the standard library. Uconses lack this integration and require a completely new set of library functions. Furthermore it must be noted that uconses are --- except if one explicitly clears the
*root-table* --- a permanent memory leak.
Yet if you are willing to accept these trade-offs, uconses offer some unique benefits:
their usage is little more expensive than a call to CONS. If you include GC time, they can even be much faster.
given enough potential for structural sharing, uconses can decrease the memory consumption of an application by orders of magnitude.
checks for structural similarity can be done in constant time. Two ucons trees are equal if and only if their roots are EQL.
Benchmarks (SBCL 1.3.20, X86-64 Intel i7-5500U CPU @ 2.40GHz):
(bench (list 1 2 3 4 5 6 7 8)) ; -> 25.77 nanoseconds (bench (ulist 1 2 3 4 5 6 7 8)) ; -> 38.18 nanoseconds
- Marco Heisig <firstname.lastname@example.org>