Awesome

Mazeppa

Supercompilation ¹ is a program transformation technique that symbolically evaluates a given program, with run-time values as unknowns. In doing so, it discovers execution patterns of the original program and synthesizes them into standalone functions; the result of supercompilation is a more efficient residual program. In terms of transformational power, supercompilation subsumes both deforestation ² and partial evaluation ³, and even exhibits certain capabilities of theorem proving.

Mazeppa is a modern supercompiler intended to be a compilation target for call-by-value functional languages. Having prior supercompilers diligently compared and revised, Mazeppa

• Provides the full set of primitive data types for efficient computation.
• Translates to native machine code through GNU C11.
• Supports manual control of function unfolding.
• Is fully transparent in terms of what decisions it takes during transformation.
• Is designed with efficiency in mind from the very beginning.

Installation

First, prepare the OCaml system on your machine:

$ bash -c "sh <(curl -fsSL https://raw.githubusercontent.com/ocaml/opam/master/shell/install.sh)"
$ opam init --auto-setup

Then install Mazeppa as an opam package:

$ opam install mazeppa

Type mazeppa --help to confirm the installation.

Alternatively, you can clone the repository and install Mazeppa manually:

$ git clone https://github.com/mazeppa-dev/mazeppa.git
$ cd mazeppa
$ ./scripts/install.sh

Flambda is a powerful program inliner and specializer for OCaml. If you build Mazeppa with an Flambda-enabled OCaml compiler, you may see much better performance. To set it up:

$ opam switch create 5.2.0+flambda ocaml-variants.5.2.0+options ocaml-option-flambda
$ eval $(opam env --switch=5.2.0+flambda)

(You may use a different version instead of 5.2.0 if you wish.)

To check if Flambda was successfully enabled, run:

$ ocamlopt -config | grep flambda

Hacking

You can play with Mazeppa without actually installing it. Having OCaml installed and the repository cloned (as above), run the following command from the root directory:

$ ./scripts/play.sh

(Graphviz is required: sudo apt install graphviz.)

This will launch Mazeppa with --inspect on playground/main.mz and visualize the process graph in target/graph.svg. The latter can be viewed in VS Code by the Svg Preview extension.

./scripts/play.sh will automatically recompile the sources in OCaml, if anything is changed.

Supercompilation by example

The best way to understand how supercompilation works is by example. Consider the following function that takes a list and computes a sum of its squared elements:

[examples/sum-squares/main.mz]

main(xs) := sum(mapSq(xs));

sum(xs) := match xs {
    Nil() -> 0i32,
    Cons(x, xs) -> +(x, sum(xs))
};

mapSq(xs) := match xs {
    Nil() -> Nil(),
    Cons(x, xs) -> Cons(*(x, x), mapSq(xs))
};

This program is written in the idiomatic, listful functional style. Every function does only one thing, and does it well. However, there is a serious problem here: mapSq essentially constructs a list that will be immediately deconstructed by sum, meaning that we 1) we need to allocate extra memory for the intermediate list, and 2) we need two passes of computation instead of one. The solution to this problem is called deforestation ², which is a special case of supercompilation.

Let us see what Mazeppa does with this program:

$ mkdir sum-squares
$ cd sum-squares
# Copy-paste the program above.
$ nano main.mz
$ mazeppa run --inspect

The --inspect flag tells Mazeppa to give a detailed report on the transformation process. The sum-squares/target/ directory will contain the following files:

target
├── graph.dot
├── nodes.json
├── output.mz
└── program.json

• graph.dot contains the complete process graph for our program. You can obtain a picture of the graph by running dot -Tsvg target/graph.dot > target/graph.svg.
• nodes.json contains the content of all nodes in the graph. Without this file, you would not be able to understand much from the graph alone.
• program.json contains the initial program in Mazeppa IR: our supercompiler works with this particular representation instead of the original program.
• output.mz contains the final residual program.

output.mz will contain the following code:

[examples/sum-squares/target/output.mz]

main(xs) := f0(xs);

f0(x0) := match x0 {
    Cons(x1, x2) -> +(*(x1, x1), f0(x2)),
    Nil() -> 0i32
};

The supercompiler has successfully merged sum and mapSq into a single function, f0! Unlike the original program, f0 works in a single pass, without having to allocate any extra memory.

How did the supercompiler got to this point? Let us see the generated process graph:

For reference, nodes.json contains the following data in JSON:

[
  [ "n0", "main(xs)" ],
  [ "n1", "sum(mapSq(xs))" ],
  [ "n2", "sum(.g1(xs))" ],
  [ "n3", "xs" ],
  [ "n4", "sum(Cons(*(.v0, .v0), mapSq(.v1)))" ],
  [ "n5", ".g0(Cons(*(.v0, .v0), mapSq(.v1)))" ],
  [ "n6", "+(*(.v0, .v0), sum(mapSq(.v1)))" ],
  [ "n7", "+(*(.v0, .v0), sum(.g1(.v1)))" ],
  [ "n8", "*(.v0, .v0)" ],
  [ "n9", ".v0" ],
  [ "n10", ".v0" ],
  [ "n11", "sum(.g1(.v1))" ],
  [ "n12", ".v1" ],
  [ "n13", "+(.v3, .v4)" ],
  [ "n14", ".v3" ],
  [ "n15", ".v4" ],
  [ "n16", "sum(Nil())" ],
  [ "n17", ".g0(Nil())" ],
  [ "n18", "0i32" ]
]

(We will not need to inspect program.json for this tiny example, but feel free to look at it: it is not too complicated.)

The supercompiler starts with node n0 containing main(xs). After two steps of function unfolding, we reach node n2 containing sum(.g1(xs)), where .g1 is the IR function that corresponds to our mapSq. At this point, we have no other choice than to analyze the call .g1(xs) by conjecturing what values xs might take at run-time. Since our mapSq only accepts the constructors Nil and Cons, it is sufficient to consider the cases xs=Cons(.v0, .v1) and xs=Nil() only.

Node n4 is what happens after we substitute Cons(.v0, .v1) for xs, where .v0 and .v1 are fresh variables. After three more function unfoldings, we arrive at n7. This is the first time we have to split the call +(*(.v0, .v0), sum(.g1(.v1))) into .v3=*(.v0, .v0) (n8) and .v4=sum(.g1(.v1)) (n11) and proceed supercompiling +(.v3, .v4) (n13); the reason for doing so is that a previous node (n2) is structurally embedded in n7, so supercompilation might otherwise continue forever. Now, what happens with sum(.g1(.v1)) (n11)? We have seen it earlier! Recall that n2 contains sum(.g1(xs)), which is just a renaming of sum(.g1(.v1)); so we decide to fold n11 into n2, because it makes no sense to supercompile the same thing twice. The other branches of n7, namely n13 and n8, are decomposed, meaning that we simply proceed transforming their arguments.

As for the other branch of n2, sum(Nil()) (n16), it is enough to merely unfold this call to 0i32 (n18).

After the process graph is completed, residualization converts it to a final program. During this stage, dynamic execution patterns become functions -- node n2 now becomes the function f0, inasmuch as some other node (n11) points to it. In any residual program, there will be exactly as many functions as there are nodes with incoming dashed lines, plus main.

In summary, supercompilation consists of 1) unfolding function definitions, 2) analyzing calls that pattern-match an unknown variable, 3) breaking down computation into smaller parts, 4) folding repeated computations, and 5) decomposing calls that cannot be unfolded, such as +(.v3, .v4) (n13) in our example. The whole supercompilation process is guaranteed to terminate, because when some computation is becoming dangerously bigger and bigger, we break it down into subproblems and solve them in isolation.

There are a plenty of other interesting examples of deforestation in the examples/ folder, including tree-like data structures. In fact, we have reimplemented all samples from the previous work on higher-order call-by-value supercompilation by Peter A. Jonsson and Johan Nordlander ^{4 5}; in all cases, Mazeppa has performed similarly or better.

Specializing the power function

Now consider another example, this time involving partial evaluation:

[examples/power-sq/main.mz]

main(a) := powerSq(a, 7u8);

powerSq(a, x) := match =(x, 0u8) {
    T() -> 1i32,
    F() -> match =(%(x, 2u8), 0u8) {
        T() -> square(powerSq(a, /(x, 2u8))),
        F() -> *(a, powerSq(a, -(x, 1u8)))
    }
};

square(a) := *(a, a);

powerSq implements the famous exponentiation-by-squaring algorithm. The original program is inefficient: it recursively examines the x parameter of powerSq, although it is perfectly known at compile-time. Running Mazeppa on main(a) will yield the following residual program:

[examples/power-sq/target/output.mz]

main(a) := let x0 := *(a, *(a, a)); *(a, *(x0, x0));

The whole powerSq function has been eliminated, thus achieving the effect of partial evaluation. (If we consider powerSq to be an interpreter for a program x and input data a, then it is the first Futamura projection: specializing an interpreter to obtain an efficient target program.) Also, notice how the supercompiler has managed to share the argument *(a, *(a, a)) twice, so that it is not recomputed each time anew. The residual program indeed reflects exponentiation by squaring.

Synthesizing the KMP algorithm

Let us go beyond deforestation and partial evaluation. Consider a function matches(p, s) of two strings, which returns T() if s contains p and F() otherwise. The naive implementation in Mazeppa would be the following, where p is specialized to "aab":

[examples/kmp-test/main.mz]

main(s) := matches(Cons('a', Cons('a', Cons('b', Nil()))), s);

matches(p, s) := go(p, s, p, s);

go(pp, ss, op, os) := match pp {
    Nil() -> T(),
    Cons(p, pp) -> goFirst(p, pp, ss, op, os)
};

goFirst(p, pp, ss, op, os) := match ss {
    Nil() -> F(),
    Cons(s, ss) -> match =(p, s) {
        T() -> go(pp, ss, op, os),
        F() -> failover(op, os)
    }
};

failover(op, os) := match os {
    Nil() -> F(),
    Cons(_s, ss) -> matches(op, ss)
};

(Here we represent strings as lists of characters for simplicity, but do not worry, Mazeppa provides built-in strings as well.)

The algorithm is correct but inefficient. Consider what happens when "aa" is successfully matched, but 'b' is not. The algorithm will start matching "aab" once again from the second character of s, although it can already be said that the second character of s is 'a'. The deterministic finite automaton built by the Knuth-Morris-Pratt algorithm (KMP) ⁶ is an alternative way to solve this problem.

By running Mazeppa on the above sample, we can obtain an efficient string matching algorithm for p="aab" that reflects KMP in action:

[examples/kmp-test/target/output.mz]

main(s) := f0(s);

f0(x0) := match x0 {
    Cons(x1, x2) -> match =(97u8, x1) {
        F() -> f1(x2),
        T() -> f2(x2)
    },
    Nil() -> F()
};

f1(x0) := f0(x0);

f2(x0) := match x0 {
    Cons(x1, x2) -> match =(97u8, x1) {
        F() -> f1(x2),
        T() -> f4(x2)
    },
    Nil() -> F()
};

f3(x0) := f2(x0);

f4(x0) := match x0 {
    Cons(x1, x2) -> match =(98u8, x1) {
        F() -> match =(97u8, x1) {
            F() -> f1(x2),
            T() -> f4(x2)
        },
        T() -> T()
    },
    Nil() -> F()
};

The naive algorithm that we wrote has been automatically transformed into a well-known efficient version! While the naive algorithm has complexity O(|p| * |s|), the specialized one is O(|s|).

Synthesizing KMP is a standard example that showcases the power of supercompilation with respect to other techniques (e.g., see ⁷ and ⁸). Obtaining KMP by partial evaluation is not possible without changing the original definition of matches ^{9 10}.

Metasystem transition

Valentin Turchin, the inventor of supercompilation, describes the concept of metasystem transition in the following way ^{11 12 13}:

Consider a system S of any kind. Suppose that there is a way to make some number of copies from it, possibly with variations. Suppose that these systems are united into a new system S' which has the systems of the S type as its subsystems, and includes also an additional mechanism which controls the behavior and production of the S-subsystems. Then we call S' a metasystem with respect to S, and the creation of S' a metasystem transition. As a result of consecutive metasystem transitions a multilevel structure of control arises, which allows complicated forms of behavior.

Thus, supercompilation can be readily seen as a metasystem transition: there is an object program in Mazeppa, and there is the Mazeppa supercompiler which controls and supervises execution of the object program. However, we can go further and perform any number of metasystem transitions within the realm of the object program itself, as the next example demonstrates.

We will be using the code from examples/lambda-calculus/. Below is a standard normalization-by-evaluation procedure for obtaining beta normal forms of untyped lambda calculus terms:

normalize(lvl, env, t) := quote(lvl, eval(env, t));

normalizeAt(lvl, env, t) := normalize(+(lvl, 1u64), Cons(vvar(lvl), env), t);

vvar(lvl) := Neutral(NVar(lvl));

eval(env, t) := match t {
    Var(idx) -> indexEnv(env, idx),
    Lam(body) -> Closure(env, body),
    Appl(m, n) ->
        let mVal := eval(env, m);
        let nVal := eval(env, n);
        match mVal {
            Closure(env, body) -> eval(Cons(nVal, env), body),
            Neutral(nt) -> Neutral(NAppl(nt, nVal))
        }
};

quote(lvl, v) := match v {
    Closure(env, body) -> Lam(normalizeAt(lvl, env, body)),
    Neutral(nt) -> quoteNeutral(lvl, nt)
};

quoteNeutral(lvl, nt) := match nt {
    NVar(var) -> Var(-(-(lvl, var), 1u64)),
    NAppl(nt, nVal) -> Appl(quoteNeutral(lvl, nt), quote(lvl, nVal))
};

indexEnv(env, idx) := match env {
    Nil() -> Panic(++("the variable is unbound: ", string(idx))),
    Cons(value, xs) -> match =(idx, 0u64) {
        T() -> value,
        F() -> indexEnv(xs, -(idx, 1u64))
    }
};

(eval/quote are sometimes called reflect/reify.)

This is essentially a big-step machine for efficient capture-avoiding substitution: instead of reconstructing terms on each beta reduction, we maintain an environment of values. eval projects a term to the "semantic domain", while quote does the opposite; normalize is simply the composition of quote and eval. To avoid bothering with fresh name generation, we put De Bruijn indices in the Var constructor and De Bruijn levels in NVar; the latter is converted into the former in quoteNeutral.

Now let us compute something with this machine:

main() := normalize(0u64, Nil(), example());

example() := Appl(Appl(mul(), two()), three());

two() := Lam(Lam(Appl(Var(1u64), Appl(Var(1u64), Var(0u64)))));

three() := Lam(Lam(Appl(Var(1u64), Appl(Var(1u64), Appl(Var(1u64),
    Var(0u64))))));

mul() := Lam(Lam(Lam(Lam(Appl(
    Appl(Var(3u64), Appl(Var(2u64), Var(1u64))),
    Var(0u64))))));

The body of main computes the normal form of the lambda term example() that multiplies Church numerals two() and three().

By supercompiling main(), we obtain the Church numeral of 6:

[examples/lambda-calculus/target/output.mz]

main() := Lam(Lam(Appl(Var(1u64), Appl(Var(1u64), Appl(Var(1u64), Appl(Var(1u64)
    , Appl(Var(1u64), Appl(Var(1u64), Var(0u64)))))))));

The lambda calculus interpreter has been completely annihilated!

In this example, we have just seen a two-level metasystem stairway (in Turchin's terminology ¹⁴): on level 0, we have the Mazeppa supercompiler transforming the object program, while on level 1, we have the object program normalizing lambda calculus terms. There can be an arbitrary number of interpretation levels, and Mazeppa can be used to collapse them all. This general behaviour of supercompilation was explored by Turchin himself in ¹ (section 7), where he was able to supercompile two interpretable programs, one Fortran-like and one in Lisp, to obtain a speedup factor of 40 in both cases.

The lambda normalizer also shows us how to incarnate higher-order functions into a first-order language. In Mazeppa, we cannot treat functions as values, but it does not mean that we cannot simulate them! By performing a metasystem transition, we can efficiently implement higher-order functions in a first-order language. Along with defunctionalization and closure conversion, this technique can be used for compilation of higher-order languages into efficient first-order code.

Related examples: imperative virtual machine, self-interpreter.

Restricting supercompilation

In retrospect, the major problem that prevented widespread adoption of supercompilation is its unpredictability -- the dark side of its power. To get a sense of what it means, consider how can we solve any SAT problem "on the fly":

[examples/sat-solver/main.mz]

main(a, b, c, d, e, f, g) := solve(formula(a, b, c, d, e, f, g));

formula(a, b, c, d, e, f, g) :=
    and(or(Var(a), or(Not(b), or(Not(c), F()))),
    and(or(Not(f), or(Var(e), or(Not(g), F()))),
    and(or(Var(e), or(Not(g), or(Var(f), F()))),
    and(or(Not(g), or(Var(c), or(Var(d), F()))),
    and(or(Var(a), or(Not(b), or(Not(c), F()))),
    and(or(Not(f), or(Not(e), or(Var(g), F()))),
    and(or(Var(a), or(Var(a), or(Var(c), F()))),
    and(or(Not(g), or(Not(d), or(Not(b), F()))),
    T()))))))));

or(x, rest) := match x {
    Var(x) -> If(x, T(), rest),
    Not(x) -> If(x, rest, T())
};

and(clause, rest) := match clause {
    If(x, m, n) -> If(x, and(m, rest), and(n, rest)),
    T() -> rest,
    F() -> F()
};

solve(formula) := match formula {
    If(x, m, n) -> analyze(x, m, n),
    T() -> T(),
    F() -> F()
};

analyze(x, m, n) := match x {
    T() -> solve(m),
    F() -> solve(n)
};

There are two things wrong with this perfectly correct code: 1) the supercompiler will expand the formula in exponential space, and 2) the supercompiler will try to solve the expanded formula in exponential time. Sometimes, we just do not want to evaluate everything at compile-time.

However, despair not: we provide a solution for this problem. Let us first consider how to postpone solving the formula until run-time. It turns out that the only thing we need to do is to annotate the function formula with @extract as follows:

@extract
formula(a, b, c, d, e, f, g) :=
    // Everything is the same.

When Mazeppa sees solve(formula(a, b, c, d, e, f, g)), it extracts the call to formula into a fresh variable .v0 and proceeds supercompiling the extracted call and solve(.v0) in isolation. The latter call will just reproduce the original SAT solver.

But supercompiling the call to formula will still result in an exponential blowup. Let us examine why this happens. Our original formula consists of calls to or and and; while or is obviously not dangerous, and propagates the rest parameter to both branches of If (the first match case) -- this is the exact place where the blowup occurs. So let us mark and with @extract as well:

@extract
and(clause, rest) := match clause {
    // Everything is the same.
};

That is it! When and is to be transformed, Mazeppa will extract the call out of its surrounding context and supercompile it in isolation. By adding two annotations at appropriate places, we have solved both the problem of code blowup and exponential running time of supercompilation. In general, whenever Mazeppa sees ctx[f(t1, ..., tN)], where f is marked @extract and ctx[.] is a non-empty surrounding context with . in a redex position, it will plug a fresh variable v into ctx and proceed transforming the following nodes separately: f(t1, ..., tN) and ctx[v].

Finally, note that @extract is only a low-level mechanism; a compiler front-end must carry out additional machinery to tell Mazeppa which functions to extract. This can be done in two ways:

• By static analysis of your object language/Mazeppa code. For example, parameter linearity analysis would mark both formula and and as extractable, leaving all other functions untouched.
• By source code annotations, in a manner similar to staged compilation.

Both methods can be combined to achieve a desired effect.

Lazy evaluation

Mazeppa employs an interesting design choice to have eager functions and lazy constructors. The following example, where magic(1u32, 1u32) generates Fibonacci numbers, was adopted from Haskell:

[examples/lazy-fibonacci/main.mz]

main() := getIt(magic(1u32, 1u32), 3u64);

magic(m, n) := match =(m, 0u32) {
    T() -> Nil(),
    F() -> Cons(m, magic(n, +(m, n)))
};

getIt(xs, n) := match xs {
    Nil() -> Panic("undefined"),
    Cons(x, xs) -> match =(n, 1u64) {
        T() -> x,
        F() -> getIt(xs, -(n, 1u64))
    }
};

If constructors were eager, magic(1u32, 1u32) would never terminate. However, Cons does not evaluate its arguments! Since getIt only consumes a finite portion of the infinite list, the program terminates and prints 2u32:

$ mazeppa eval
2u32

Lazy constructors enable effortless deforestation, as discussed below.

Mazeppa as a library

After installing Mazeppa via opam or ./scripts/install.sh, it is available as an OCaml library!

Set up a new Dune project as follows:

$ dune init project my_compiler

Add mazeppa as a third-party library into your bin/dune:

(executable
 (public_name my_compiler)
 (name main)
 (libraries my_compiler mazeppa))

Paste the following code into bin/main.ml (this is examples/sum-squares/main.mz encoded in OCaml):

open Mazeppa

let input : Raw_program.t =
    let sym = Symbol.of_string in
    let open Raw_term in
    let open Checked_oint in
    [ [], sym "main", [ sym "xs" ], call ("sum", [ call ("mapSq", [ var "xs" ]) ])
    ; ( []
      , sym "sum"
      , [ sym "xs" ]
      , Match
          ( var "xs"
          , [ (sym "Nil", []), int (I32 (I32.of_int_exn 0))
            ; ( (sym "Cons", [ sym "x"; sym "xs" ])
              , call ("+", [ var "x"; call ("sum", [ var "xs" ]) ]) )
            ] ) )
    ; ( []
      , sym "mapSq"
      , [ sym "xs" ]
      , Match
          ( var "xs"
          , [ (sym "Nil", []), call ("Nil", [])
            ; ( (sym "Cons", [ sym "x"; sym "xs" ])
              , call
                  ( "Cons"
                  , [ call ("*", [ var "x"; var "x" ]); call ("mapSq", [ var "xs" ]) ] ) )
            ] ) )
    ]
;;

let () =
    try
      let output = Mazeppa.supercompile input in
      Printf.printf "%s\n" (Raw_program.show output)
    with
    | Mazeppa.Panic msg ->
      Printf.eprintf "Something went wrong: %s\n" msg;
      exit 1
;;

Run dune exec my_compiler to see the desired residual program:

[([], "main", ["xs"], (Raw_term.Call ("f0", [(Raw_term.Var "xs")])));
  ([], "f0", ["x0"],
   (Raw_term.Match ((Raw_term.Var "x0"),
      [(("Cons", ["x1"; "x2"]),
        (Raw_term.Call ("+",
           [(Raw_term.Call ("*", [(Raw_term.Var "x1"); (Raw_term.Var "x1")]));
             (Raw_term.Call ("f0", [(Raw_term.Var "x2")]))]
           )));
        (("Nil", []), (Raw_term.Const (Const.Int (Checked_oint.I32 0))))]
      )))
  ]

You can call Mazeppa as many times as you want, including in parallel. Note that we expose a limited interface to the supercompiler; in particular, there is no way to inspect what it does in the process (i.e., --inspect).

Besides supercompilation, we also provide a built-in evaluator:

val eval : Raw_program.t -> Raw_term.t

It can only be called on programs whose main functions do not accept parameters. Unlike supercompile, it produces an evaluated term of type Raw_term.t and can possibly diverge.

See other API functions and their documentation in lib/mazeppa.mli.

Translation to C

Suppose that main.mz contains a slightly modified version of the lazy Fibonacci example:

main(n) := getIt(magic(1u32, 1u32), n);

magic(m, n) := match =(m, 0u32) {
    T() -> Nil(),
    F() -> Cons(m, magic(n, +(m, n)))
};

getIt(xs, n) := match xs {
    Nil() -> Panic("undefined"),
    Cons(x, xs) -> match =(n, 1u64) {
        T() -> x,
        F() -> getIt(xs, -(n, 1u64))
    }
};

The following command translates it to C11 with GNU extensions (i.e., -std=gnu11):

$ cat main.mz | mazeppa translate --language C --entry fib

#include "mazeppa.h"

MZ_ENUM_USER_TAGS(op_Cons, op_Nil);

static mz_Value op_main(mz_ArgsPtr args);

static mz_Value op_magic(mz_ArgsPtr args);

static mz_Value op_getIt(mz_ArgsPtr args);

static mz_Value thunk_0(mz_EnvPtr env) {
    mz_Value var_m = (env)[0];
    mz_Value var_n = (env)[1];
    return ({
    struct mz_value args[2];
    (args)[0] = var_n;
    (args)[1] = MZ_OP2(var_m, add, var_n);
    op_magic(args);
    });
}

static mz_Value op_main(mz_ArgsPtr args) {
    mz_Value var_n = (args)[0];
    return ({
    struct mz_value args[2];
    (args)[0] = ({
    struct mz_value args[2];
    (args)[0] = MZ_INT(U, 32, UINT32_C(1));
    (args)[1] = MZ_INT(U, 32, UINT32_C(1));
    op_magic(args);
    });
    (args)[1] = var_n;
    op_getIt(args);
    });
}

static mz_Value op_magic(mz_ArgsPtr args) {
    mz_Value var_m = (args)[0];
    mz_Value var_n = (args)[1];
    return ({
    struct mz_value tmp = MZ_OP2(var_m, equal, MZ_INT(U, 32, UINT32_C(0)));
    switch((tmp).tag) {
    case op_T: {
    tmp = MZ_EMPTY_DATA(op_Nil);
    break;
    }
    case op_F: {
    tmp = MZ_DATA(op_Cons, 2, MZ_SIMPLE_THUNK(var_m), MZ_THUNK(thunk_0, 2, var_m, var_n));
    break;
    }
    default: MZ_UNEXPECTED_TAG((tmp).tag);
    }
    tmp;
    });
}

static mz_Value op_getIt(mz_ArgsPtr args) {
    mz_Value var_xs = (args)[0];
    mz_Value var_n = (args)[1];
    return ({
    struct mz_value tmp = var_xs;
    switch((tmp).tag) {
    case op_Nil: {
    tmp = mz_panic(MZ_STRING("undefined"));
    break;
    }
    case op_Cons: {
    mz_Value var_x = ((tmp).payload)[0];
    mz_Value var_xs$ = ((tmp).payload)[1];
    tmp = ({
    struct mz_value tmp = MZ_OP2(var_n, equal, MZ_INT(U, 64, UINT64_C(1)));
    switch((tmp).tag) {
    case op_T: {
    tmp = mz_force(var_x);
    break;
    }
    case op_F: {
    tmp = ({
    struct mz_value args[2];
    (args)[0] = mz_force(var_xs$);
    (args)[1] = MZ_OP2(var_n, sub, MZ_INT(U, 64, UINT64_C(1)));
    op_getIt(args);
    });
    break;
    }
    default: MZ_UNEXPECTED_TAG((tmp).tag);
    }
    tmp;
    });
    break;
    }
    default: MZ_UNEXPECTED_TAG((tmp).tag);
    }
    tmp;
    });
}

extern mz_Value fib(mz_Value var_n) {
    return MZ_CALL_MAIN(var_n);
}

The translate command requires both --language , which is the target language for translation, and --entry, which is the name of an external symbol that will correspond to your main function. The input Mazeppa program comes from stdin (cat main.mz in our example); the output GNU11 program is written to stdout.

Let us advance further and compile the output program to an object file. First, copy c/deps/sds.c, c/deps/sds.h, and c/deps/sdsalloc.h to your current directory; second, install Boehm GC on your computer:

$ sudo apt install libgc-dev -y

then execute the following command:

$ cat main.mz \
    | mazeppa translate --language C --entry fib --dump-header-to . \
    | gcc -c -o program.o -std=gnu11 -xc -

The --dump-header-to option writes the content of mazeppa.h to a specified location; this is needed for the output program to compile. The gcc command accepts the output program from stdin and produces program.o.

Now what is left is to actually invoke the generated fib function. Create main.c with the following content:

#include "mazeppa.h"

mz_Value fib(mz_Value n);

int main(void) {
    // Always initialize Boehm GC before invoking Mazeppa code.
    GC_INIT();
    mz_Value v = fib(MZ_INT(U, 64, 10));
    printf("fib(10) = %" PRIu32 "\n", MZ_GET(U32, v));
}

This "driver" program just invokes fib with a Mazeppa integer (MZ_INT) and prints the result. You can use any functionality from mazeppa.h, provided that it is not prefixed with mz_priv_ or MZ_PRIV_.

To bring all the pieces together:

$ gcc main.c program.o sds.c -lgc -std=gnu11

./a.out prints fib(10) = 55 and exits, as expected.

Usage considerations

• Follow the instructions from the installation section to build Mazeppa with an Flambda-enabled OCaml compiler.
• Forget about --inspect in a real environment: use it only for debugging purposes.
• Refrain from CPU-intensive computations: termination checking makes them extremely slow. Use @extract (as shown above) to control whether to continue supercompilation or extract the redex.
• Supercompiling a whole user program is not a good idea. Instead, consider supercompiling separate modules and then linking them together (the exact way you do so depends on your situation).
  • Although Mazeppa is itself not parallel, supercompiling separate modules can be done in parallel.
• It is much better if recursive functions reduce at least one argument structurally, just as in total functional programming. Otherwise, termination checking might not work well.
  • No integer is structurally smaller than another one: iteration can happen in any order.
  • A string s1 is smaller than s2 iff s2 contains all the characters of s1, preserving the order; for example, "otus" is smaller than "octopus" but "octopusx" is not.

Technical decisions

Mazeppa employs several interesting design choices (ranked by importance):

• First-order code. Max Bolingbroke and Simon Peyton Jones ¹⁵ report that for one particular example, their higher-order supercompiler for a subset of Haskell spent 42% of execution time on managing names and renaming. While it is true that simplistic evaluation models, such as normalization of lambda terms, permit us to avoid significant overhead of capture avoidance, supercompilation is more complicated. For example, besides doing symbolic computation, supercompilation needs to analyze previously computed results to make informed decisions about further transformation: consider term instance tests, homeomorphic embedding tests, most specific generalizations, etc. Introducing higher-order functions inevitably complicates all these analyses, making supercompilation slower, more memory-consuming, and harder to reason about. In Mazeppa, we stick with the philosophy of gradual improvements: instead of trying to handle many fancy features at the same time, we 1) fix the core language for convenient manipulation by a machine, 2) perform as many metasystem transitions as necessary to make the core language better for human.

• Lazy constructors. It is a well-known observation that call-by-value languages are hard for proper deforestation. It is still possible to deforest them, but not without additional analysis ^{4 5}. However, if constructors are lazy (i.e., they do not evaluate their arguments), deforestation just works. Turchin made it work by normal-order transformation of a call-by-value language, but the result is that residual code may terminate more often. In Mazeppa, we have call-by-value functions and call-by-name (call-by-need) constructors, which 1) makes deforestation possible and 2) preserves the original semantics of code.

  • Incidentally, lazy constructors are also adopted by eager functional languages outside of supercompilation. See "Why is Idris 2 so much faster than Idris 1?" and "CONS Should Not Evaluate its Arguments" ¹⁶.

• Term-free process graphs. In Mazeppa, process graphs do not contain any references to terms: residualization can work without them. As a result, the garbage collector can deallocate terms that were used during construction of a subgraph. In addition ot that, this policy has several other important advantages: 1) the graph is still there for inspection with --inspect, 2) when it is drawn, it only reveals information about the decisions the supercompiler has taken, which makes it much easier to look at. Several existing supercompilers refrain from proper process graphs (e.g., Neil Mitchell's Supero ¹⁷ and the aforementioned ¹⁵), but as a result, 1) they are less capable of inspection by a user, 2) the algorithms become cluttered with code generation details.

• Two-dimensional configuration analysis. Usually a supercompiler keeps a "history" of a subset of all ancestors while transforming a node; if this node is "close enough" to one of its ancestors, it is time to break the term into smaller parts to guarantee termination. In Mazeppa, we keep two separate data structures instead: the one containing a subset of node's ancestors and the one containing a subset of fully transformed nodes. The former data structure is used to guarantee termination (as usual), while the latter is used to enhance sharing of functions in residual code. Specifically, if the current node (of special kind) is a renaming of some previously transformed node, we fold the current node into this previous node. This way, Mazeppa performs both vertical and horizontal analysis of configurations, which makes residual code more compact and supercompilation more efficient.

• Function productivity analysis. When an inner function call pattern-matches an unknown value, two potentially dangerous things happen: 1) supercompilation reproduces the structure of the pattern-matching function, 2) supercompilation pushes the whole surrounding context to all branches of this function. Without further control, this situation can lead to significant explosion in code size, sometimes even causing a supercompiler not to terminate in a reasonable amount of time. To ameliorate, Mazeppa duplicates the context iff this inner call produces a definite top-level value from at least one exit point, because if it does, great chances that this value can be deconstructed by subsequent pattern matching. Otherwise, Mazeppa extracts the inner call and transforms it in isolation (just as if it was marked with @extract). Furthermore, if the context is actually duplicated, the following rule applies to all branches: if the branch produces a definite top-level value from all exit points, it is transformed in the context as usual; otherwise, the branch is extracted from the context and transformed in isolation. In practice, this analysis prevents a huge amount of unneeded specializations, thereby compactifying residual program size and making supercompilation much more tractable.

  • We perform productivity analysis strictly before supercompilation. The algorithm launches global analysis for all f- and g-rules present in Mazeppa IR, and has linear time complexity.
  • An emergent effect is that after extracting these pattern-matching calls, horizontal configuration analysis becomes able to fold more repeated computations.

• Smart histories. Instead of blindly comparing a current node with all its ancestors, we employ a more fine-grained control, that is: 1) global nodes (the ones that analyze an unknown variable) are compared with global nodes only, 2) local nodes (the ones that reduce linearly in a single step) are compared with local nodes only up to the latest global node, but not including it, and 3) trivial nodes (the ones that break down terms into smaller components) are not compared with anything else. Besides a more economic approach to termination checking, this scheme allows Mazeppa to discover more optimization opportunities; see ¹⁸, sections 4.6 and 4.7. Termination of supercompilation is guaranteed by the fact that homeomorphic embedding is still tested on all potentially infinite subsequences of global and local terms (there cannot exist an infinite sequence of trivial terms only).

• Redex signatures. The whistle is only tested on terms with equal redex signatures. A redex signature is a pair of 1) a function's symbol and 2) a list of argument value categories. A value category is a sort of metainformation about an argument, which can be either 1) VConst for such values as 42i32 or "hello world", 2) VNeutral for such values as x or +(x, x), or 3) VCCall(c) for constructor calls such as Foo(...). The whistle is still tested on all potentially infinite sequences of terms, because any infinite sequence must contain at least one infinite subsequence of terms with the same redex signature. This strategy allows Mazeppa to avoid over-generalization in certain cases, as the following two examples demonstrate:

  • Redex operators: during execution, node f(A()) is homeomorphically embedded into f(g(B(A()))); however, since the redex operators are different (f and g), Mazeppa continues reduction and reaches the ideal result.
  • Redex signatures: although f(A(Good())) is embedded into f(f(f(B(A(Good()))))) and the redex operator is f in both cases, Mazeppa does not over-generalize because these two terms have different value categories in their redex signatures: in the first case, f is called on A(...), while in the second case, it is called on B(...). Just as in the previous example, Mazeppa reaches the ideal result of supercompilation solely by reduction.

• Optimized homeomorphic embedding. Over the decades, homeomorphic embedding has gained reputation for being an excellent method for online termination checking ¹⁹. Unfortunately, mainly due to the non-linear control flow (when both diving and coupling apply), it can be inexcusably expensive to compute. What is even worse, it is re-executed for all qualifying parent nodes whenever a new term is added to the history, which progressively slows down supercompilation as the history grows: going as far as eating most of the supercompilation time! To cope with that, we maintain two separate caches:

  • Global cache: an ephemeron from term addresses to integers denoting their sizes. Consider two terms, t1 and t2; if the size of t1 is equal to the size of t2, t1 can only be embedded into t2 by coupling. On the other hand, if the size of t1 is smaller or equal to the size of t2, both diving and coupling may take place. Otherwise, t1 is a bigger term which cannot be embedded into t2 by no means. Whenever there is a choice to dive or/and couple, we query the cache and decide what to do; this technique can rule out many redundant control flow paths. The idea is borrowed from ²⁰, section 6.3.
    • Note that deriving the value address is a constant-time perfect hash function, so the cache penalty should be pretty low. Since it is ephemeral, it also does not induce memory leaks.
  • Local cache: a hash table from addresses of terms (t1, t2) to booleans denoting results of their comparison. This cache makes homeomorphic embedding to have O(size(t1) * size(t2)) worst-case time complexity by not recomputing physically equal terms. Unlike the global cache, the local cache only exists during a single comparison.

• Hash consing. The homeomorphic embedding caches described above depend on how many terms are shared. We therefore employ hash consing while unfolding function bodies: if some new term t is structurally equal to some existing term s, the latter is reused. To avoid memory leaks, we employ a global ephemeron holding weak pointers to terms. Besides improved supercompilation times (due to memoization), hash consing also reduces memory consumption -- see https://github.com/mazeppa-dev/mazeppa/issues/4#issuecomment-2282772940.

• Normalization during unfolding. When a function call is unfolded, we substitute the parameters and normalize the body as much as possible (i.e., without further unfoldings, to guarantee termination). To see why, consider the factorial function f(n); with simple unfolding, we would trap in an unpleasant situation where f(1u32) is embedded into *(1u32, f(-(1u32, 1u32))), causing over-generalization. In reality, Mazeppa would unfold f(1u32) to *(1u32, f(0u32)), making the latter a candidate for further unfolding. This approach was suggested in ²⁰, section 4.5. Its other merits are: 1) less work for future driving steps, 2) less "banal" computation in process graphs, 3) reduced amount of expensive homeomorphic embedding tests.

  • Besides folding constants, we also perform algebraic simplification such as +(t, 0), +(0, t) -> t, -(t, 0) -> t, *(t, 0), *(0, t) -> 0, etc. These can be seen as axioms for built-in types.
  • After implementing redex signatures, *(1u32, f(-(1u32, 1u32))) is no longer checked against f(1u32) because of different redex operators. However, other advantages of normalization still hold.

• Implementation in OCaml. Mazeppa is implemented using a combination of functional and imperative programming styles, which is very natural to do in OCaml. Exceptions are used not only for "exceptional" situations, mutability inside functions is common. Although we do not have a similar supercompiler written in e.g. Haskell or Rust for comparison, we believe that it is OCaml that gave us a working implementation without having to quarrel with the language and never finishing work.

While most of the above is not particularly novel, we believe that the combination of these features makes Mazeppa a more practical alternative than its predecessors.

Further research

• Can supercompilation be used to erase unnecessary information from dependently typed programs ²¹?
• Is it possible to devise heuristics that would guarantee predictability of supercompilation? (E.g., by dynamically marking misbehaving functions as extractable.)
• What if we partially evaluate the supercompiler based on a set of function definitions (e.g., some fixed interpreter)? This could make supercompilation significantly more efficient.
• Equality indices ²² can enhance dynamic sharing of arguments. Would it be beneficial to implement them in Mazeppa?
• Suppose that there exists a dirty language L, a pure interpreter I for L, and a fixed program P in L. By supercompiling I(P, data), where data is unknown, we should be able to automatically purify the program P! Likewise, we should be able to purify not only all programs in L, but also all dirty languages with interpretive definitions in Mazeppa.

Language definition

Lexical structure

A symbol <SYMBOL> is a sequence of letters (a, ..., z and A, ..., Z) and digits (0, ..., 9), followed by an optional question mark (?), followed by an optional sequence of ' characters. The underscore character (_) may be the first character of a symbol, which may informally indicate that the value or function being defined is not used; otherwise, the first character must be a letter. The following sequences of characters are also permitted as symbols: ~, #, +, -, *, /, %, |, &, ^, <<, >>, =, !=, >, >=, <, <=, ++. The following are reserved words that may not be used as symbols: match, let.

There are four classes of unsigned integer constants:

• Binary: 0b (0B) followed by a non-empty sequence of binary digits 0 and 1.
• Octal: 0o (0O) followed by a non-empty sequence of octal digits 0, ..., 7.
• Decimal: a non-empty sequence of decimal digits 0, ..., 9.
• Hexadecimal: 0x (0X) followed by a non-empty sequence of decimal digits 0, ..., 9 and letters a, ..., f (A, ..., F).

Notes:

• For convenience of reading, each unsigned integer constant may contain underscore characters (_) except for the first position in the sequence of digits.
• For each unsigned integer constant, there is a negated integer constant formed by the negation character (-) placed right before the sequence of digits and underscore characters.
• For each unsigned and negated integer constant, there is a typed integer constant <INT> produced by appending an integer type <INT-TY> (u8, u16, u32, u64, u128, i8, i16, i32, i64, i128) right after the original integer constant. For example, the constants 123i8, 123u16, and 123i32 all belong to the set <INT>.

A string constant <STRING> is a sequence, between double quotes ("), of zero or more printable characters (we refer to printable characters as those numbered 33-126 in the ASCII character set), spaces, or string escape sequences:

where h is either 0, ..., 9 or a, ..., f or A, ..., F.

A character constant <CHAR> is either a sole character enclosed in single quotes (') or a character escape sequence enclosed in single quotes. The character escape sequence is the same as for strings, except that \" is replaced by \'.

There are no other constants in Mazeppa.

A comment <COMMENT> is any sequence of characters after //, which is terminated by a newline character. (We only allow single-line comments for simplicity.)

Surface syntax

The entry point <program> is defined by the following rules:

• <def-attr-list> <SYMBOL> ( <SYMBOL>, ..., <SYMBOL> ) := <term> ; <program>
• <COMMENT> <program>
• (Empty string.)

where <def-attr-list> is a whitespace-separated sequence of function attributes (the same attribute can occur multiple times). Right now, the only allowed function attribute is @extract.

<term> is defined as follows:

• <SYMBOL> (a variable)
• <const> (a constant)
• <SYMBOL> ( <term>, ..., <term> ) (a function call)
• match <term> { <match-case>, ..., <match-case> } (pattern matching)
• let <SYMBOL> := <term> ; <term> (a let-binding)
• let <pattern> := <term> ; <term> (a pattern let-binding)
• <COMMENT> <term> (a comment)

The rest of the auxiliary rules are:

<const>:

• Either <INT> or <STRING> or <CHAR>.

<match-case>:

• <pattern> -> <term>

<pattern>:

• <SYMBOL> ( <SYMBOL>, ..., <SYMBOL> ).

In Mazeppa, primitive operations employ the same syntax as that of ordinary function calls. To distinguish between the two, we define <op1> and <op2> to be the following sets of symbols:

• <op1> is one of ~, #, length, string, <INT-TY>.
• <op2> is one of +, -, *, /, %, |, &, ^, <<, >>, =, !=, >, >=, <, <=, ++, get.

Furthermore, <op2> has the following subclasses:

• <arith-op2> is one of +, -, *, /, %, |, &, ^, <<, >>.
• <cmp-op2> is one of =, !=, >, >=, <, <=.

Restrictions

• Per-program: 1) No symbol can be called with a different number of arguments. 2) No constructor symbol X can be called with N arguments but matched with M arguments, where N != M. 3) If some function is defined with N parameters, it must be called with exactly N arguments. 4) No two functions can define the same symbol.
• Per-function: 1) No function can redefine a primitive operator <op1> or <op2>. 2) A function must define a symbol starting with a lowercase letter. 3) No duplicate symbols can occur among function parameters. 4) Every free variable inside a function body must be bound by a corresponding parameter in the function definition.
• Per-term: 1) The sequence of cases in match { ... } must not be empty. 2) No duplicate constructors can occur among case patterns in match { ... }. 3) No duplicate symbols can occur among pattern parameters C(x1, ..., xN). 4) No let-binding can bind <op1> or <op2>. 5) Panic must be called with only one argument; T and F with zero arguments.

If a program, function, or term conforms to these restrictions, we call it well-formed.

Desugaring

where ASCII(c) is an appropriate u8 integer constant, according to the ASCII table; for example, ASCII('a') is 97u8.

Evaluation

Suppose that t is a well-formed term closed under environment env (defined below) and program is a well-formed program. Then the evaluation of t is governed by the following big-step environment machine:

• eval(env, x) = eval({ }, env(x))
• eval(env, const) = const, where const is in <const>.
• eval(env, f(t1, ..., tN)) =
  • t1Val ^= eval(env, t1)
  • ...
  • tNVal ^= eval(env, tN)
  • eval({ x1 -> t1Val, ..., xN -> tNVal }, body), where f(x1, ..., xN) := body; is in program.
• eval(env, C(t1, ..., tN)) = C(t1[env], ..., tN[env]), where C starts with an uppercase letter.
• eval(env, op(t)) =
  • tVal ^= eval(env, t)
  • evalOp1(op, tVal), where op is in <op1>.
• eval(env, op(t1, t2)) =
  • t1Val ^= eval(env, t1)
  • t2Val ^= eval(env, t2)
  • evalOp2(t1Val, op, t2Val), where op is in <op2>.
• eval(env, match t { p1 -> t1, ..., pN -> tN }) =
  • Ci(s1, ..., sN) ^= eval(env, t)
  • eval(env', tI), where
    • Ci(x1, ..., xN) -> tI is among the rules in match t { ... }, and
    • env' is env[x1 -> s1, ..., xN -> sN].
• eval(env, let x := t; u) =
  • tVal ^= eval(env, t)
  • eval(env[x -> tVal], u)

Notation:

• env is a total environment over t, whose general form is { x1 -> t1, ..., xN -> tN }. Each tI term must be closed and well-formed; this property is preserved by eval.
• env(x) is tI, where x -> tI is in env.
• env[x1 -> t1, ..., xN -> tN] is the environment env extended with new bindings x1 -> t1, ..., xN -> tN. If some xI is already bound in env, it is rebound.
• t[env] denotes a simultaneous substitution of all free variables in t by their bound values in env.
• tVal ^= eval(env, t) evaluates t under env; then:
  • If it is Panic(t'), the result of the whole evaluation rule is Panic(t').
  • Otherwise, tVal is available for the next clauses.

(Note that eval is a partial function, so evaluation of t can "get stuck" without a superimposed type system.)

In what follows, 1) signed integers are represented in two's complement notation, 2) panic denotes Panic(s), where s is some (possibly varying) implementation-defined string constant.

evalOp1 takes care of the unary operators for primitive types (x is in <INT>, s is in <STRING>):

• evalOp1(~, x) is the bitwise negation of x of the same type.
• evalOp1(string, x) is the string representation of x (in decimal).
• evalOp1(ty, x), where ty is in <INT-TY>, is x converted to an integer of type ty.
  • If x is not in the range of ty, the result is panic.
• evalOp1(#, x), where x is a u8 integer, is a string containing only the ASCII character of x.
  • If x is not printable, the result takes the form "\xhh".
• evalOp1(length, s) is a u64 integer denoting the length of s.
• evalOp1(string, s) is s.

Likewise, evalOp2 takes care of the binary operators for primitive types:

• evalOp2(x, op, y), where x and y have the same integer type and op is in <arith-op2>, performs a corresponding arithmetic operation on x and y, yielding a value of the same type as that of x and y.
  • If the value is not representable in that type, the result is panic.
• evalOp2(x, op, y), where x and y have the same integer type and op is in <cmp-op2>, performs a corresponding comparison operation on x and y, yielding either T() or F().
• evalOp2(s1, op, s2), where s1 and s2 are strings and op is in <cmp-op2>, performs a corresponding lexicographical comparison on s1 and s2, yielding either T() or F().
• evalOp2(s1, ++, s2) is the concatenation of s1 and s2.
• evalOp2(s, get, idx), where idx is a u64 integer, is the idx-th character (of type u8) of s.
  • If idx is out of bounds, the result is panic.

The definition of eval is now complete.

Release procedure

1. Update the version field in dune-project and bin/main.ml.
2. Type dune build to generate mazeppa.opam.
3. Update CHANGELOG.md.
4. Release the project in GitHub Releases.

FAQ

Is Mazeppa production-ready?

Not yet, we need to battle-test Mazeppa on some actual programming language. Our long-term goal is to find suitable heuristics to profitably supercompile any source file under 10'000 LoC (in Mazeppa).

How can I execute programs in Mazeppa?

For debugging and other purposes, we provide a built-in definitional interpreter that can execute Mazeppa programs. You can launch it by typing mazeppa eval (make sure that your main function does not accept parameters). For the purpose of real execution, we recommend translating Mazeppa to C and then compiling C to machine code, as shown above.

How can I perform I/O in Mazeppa?

Since Mazeppa is a purely functional language, the only way to implement I/O is as in Haskell ²³: having a pure program that performs computation and a dirty runtime that performs side effects issued by the program. There are no plans to introduce direct I/O into Mazeppa: it will only make everything more complicated.

Will Mazeppa have a type system?

No, we do not think that a type system is necessary at this point. It is the responsibility of a front-end compiler to ensure that programs do not "go wrong".

Can I use Mazeppa for theorem proving?

The more we make supercompilation predictable, the less it is capable of theorem proving. For those interested in program analysis rather than optimization, we suggest looking into distillation ²⁴.

Where can I learn more about supercompilation?

For the English audience, the following paper presents a decent introduction into supercompilation:

• "Supercompilation: Ideas and Methods" ²⁵

However, the following papers in Russian describe a supercompilation model that is closer the majority of existing supercompilers, including Mazeppa:

• "Supercompilation: main principles and basic concepts" ²⁶
• "Supercompilation: homeomorphic embedding, call-by-name, partial evaluation" ²⁷

Mazeppa itself is inspired by this excellent paper (in English):

• "A Roadmap to Metacomputation by Supercompilation" ²⁸

Finally, the international META workshops are great collections of articles about supercompilation and adjacent fields:

• META 2008: http://meta2008.pereslavl.ru/
• META 2010: http://meta2010.pereslavl.ru/
• META 2012: http://meta2012.pereslavl.ru/
• META 2014: http://meta2014.pereslavl.ru/
• META 2016: http://meta2016.pereslavl.ru/

Can supercompilation be even more powerful?

Several approaches can lead to superlinear speedup of non-esoteric programs by supercompilation:

• Jungle driving (appeared 2001) ²⁹ uses jungle evaluation ³⁰ as an advanced driving strategy for first-order positive supercompilation. The key idea behind jungle evaluation is to share as much computation as theoretically possible by reducing graphs (jungles) instead of terms. Consider that the structure of a supercompiled program reflects the way it will be executed at run-time; in turn, the strategy of driving is what affects the structure. Driving jungles instead of terms can therefore lead to exponentially more efficient residual programs, but the (seemingly inevitable) cost of managing jungles remains at compile-time!
• Distillation (appeared 2007) ^{31 32 33 34} is an upgraded version of higher-order positive supercompilation. The main difference is that generalization and folding are performed with respect to graphs instead of terms, thus allowing the distiller to make more insightful decisions about program transformation.
  • Furthermore, distillation can be employed to define a hierarchy of program transformers ³⁵, where each higher-level transformer is progressively more powerful than lower-level ones.
• Higher-level supercompilation ³⁶ utilizes lower-level supercompilers to discover lemmas about term equivalences, which are then used by higher-level supercompilers for insightful term rewriting. However, to the best of our knowledge, no fully automatic procedure for discovering (and applying) lemmas in the infinite search space has been proposed yet.

None of the above is planned to be implemented in Mazeppa, because 1) we think that writing asymptotically good programs is the responsibility of the programmer, not the optimizer, and 2) predictability of supercompilation is of greater importance to us. However, for those who are interested in this topic, the references may be helpful.

How do I contribute?

Just fork the repository, work in your own branch, and submit a pull request. Prefer rebasing when introducing changes to keep the commit history as clean as possible.

References