Thoughts about supercompilation and optimal reduction

[Tima Kinsart]

I've been interested in these two topics lately. At first sight, they seem to be very different (and they are!), but I've started coming to a conjecture that they both can be complementary.

Supercompilation is a program transformation technique that is usually run at compile-time ¹. It builds a graph (a (partial) process tree) approximating all possible execution states for a particular program, optimizes this graph by removing computational redundancy, and outputs a residual program that has a simpler and (hopefully) more efficient structure. Notwithstanding that supercompilation mimics run-time execution, it can be made terminating even for a Turing-complete language using folding and generalization of terms.

Optimal reduction is a technique to reduce lambda terms to their beta normal form. It does so by building a graph for an initial term (an interaction net), then transforming the graph with a fixed set of local rewrite rules. When the job is done, it outputs a residual program representing a beta normal form of the initial term, which has a simpler structure. Due to the halting problem, optimal reduction is not guaranteed to terminate on every input term.

A supercompiler can be understood as a program that inlines every single thing. An optimal evaluator can be understood as a program that shares every single thing. So, in principle, these two techniques are diverging in this matter: while a supercompiler attempts to remove sharing, an optimal evaluator does the opposite.

One of the most dangerous problems with supercompilation is that, due to a disrespectful attitude to sharing, it can output a term that is much larger than the input term. On the other hand, this is absolutely not a problem with optimal reduction, because a very compact interaction net can be residualized into a very large term. Therefore, one potential approach to solving the supercompilation code bloat problem would be to escape the syntactical universe of input terms and instead manipulate them in a somewhat more intricate manner that permits an all-encompassing notion of sharing. In a sense, Lamping's optimal reduction already provides such a mechanism.

It would be pretty interesting to see how optimal reduction's techniques can be applied to supercompilation. Because of sharing, a residual program can be made "as small as possible" using some sort of a graph representation. When it is given run-time data to act upon, it can be evaluated by some sort of a graph machine.

Anyhow, if you find the discussion interesting, please share (punch unintended!) any thoughts about it.

---

By the way, the README section "Is HVM's optimality only relevant for weird, academic λ-encoded terms?" demonstrates a truly interesting example of annihilating an intermediate list for the map f . map g term. In "Rethinking Supercompilation", Neil Mitchell demonstrates the same term encoded as follows (section 2):

root g f x = map g (map f x)

map f [] = []
map f (x : xs) = f x : map f xs

Which the suggested supercompiler is able to optimize into a single-pass function (section 2.3.2):

root g f x = case x of
    [] -> []
    z : zs -> g (f z) : root g f zs

Which is computationally identical to HVM's produced result map (f . g).

I believe that even a simple supercompiler would be also able to optimize your next example foldr (.) id funcs. So, in a sense, HVM and supercompilers have a common set of optimizations they're able to perform.

Footnotes

1. In theory, it can be executed at run-time, but I haven't seen a single experiment. ↩