pcf

A one file compiler for PCF.

PCF is a small programming language with higher order functions, natural numbers, and recursion. It is statically tpyed and turing complete (general recursion and all that). This compiler transformers a PCF expression into a file of C code that when run outputs the answer. It is mostly intended as a demonstration of how to write such a compiler. The curious reader should look at the writeup.

pcf

A one file compiler for PCF to C. It's currently about 275 lines of compiler and 75 lines of extremely boring instances. The compiler is fully explained in this blog post.

What's PCF

PCF is a tiny typed, higher-order functional language. It has 3 main constructs,

1. Natural Numbers

   In PCF there are two constants for natural numbers. Zero and Suc. Zero is pretty self explanatory. Suc e is the successor of a natural number, it's 1 + e in other languages. Finally, when given a natural number you can pattern match on it with ifz.

    ifz e {
       Zero  => ...
     | Suc x => ...
    }
   

   Here the first branch runs if e evaluates to zero and the second branch is run if e evaluates to Suc .... x is bound to the predecessor of e in the successor case.

2. Functions

   PCF has functions. They can close over variables and are higher order. Their pretty much what you would expect from a functional language. The relevant words here are Lam and App. Note that functions must be annotated with their arguments type.

3. General Recursion

   PCF has general recursion (and is thus Turing complete). It provides it in a slightly different way than you might be used to. In PCF you have the expression fix x : t in ... and in the ...x would be bound. The intuition here is that x stands for the whole fix x : t in ... expression. If you're a Haskell person you can just desugar this to fix $ \x : t -> ....

Example

For a quick example of how this all hangs together, here is how you would define plus in PCF

    plus =
      fix rec : nat -> nat -> nat in
        λ m : nat.
        λ n : nat.
          ifz m {
             Zero  => n
           | Suc x => Suc (rec x n)
          }

For this library we'd write this AST as

    let lam x e = Lam Nat $ abstract1 x e
        fix x e = Fix (Arr Nat (Arr Nat Nat)) $ abstract1 x e
        ifz i t x e = Ifz i t (abstract1 x e)
        plus = fix 1 $ lam 2 $ lam 3 $
                 ifz (V 2)
                     (V 3)
                     4 (Suc (App (V 1) (V 4) `App` (V 3)))
    in App (App plus (Suc Zero)) (Suc Zero)

We can then chuck this into the compiler and it will spit out the following C code

    tagged_ptr _21(tagged_ptr * _30)
    {
        tagged_ptr _31 = dec(_30[1]);
        tagged_ptr _35 = EMPTY;
        if (isZero(_30[1]))
        {
            _35 = _30[2];
        }
        else
        {
            tagged_ptr _32 = apply(_30[0], _31);
            tagged_ptr _33 = apply(_32, _30[2]);
            tagged_ptr _34 = inc(_33);
            _35 = _34;
        }
        return _35;
    }
    tagged_ptr _18(tagged_ptr * _36)
    {
        tagged_ptr _37 = mkClos(_21, 2, _36[0], _36[1]);
        return _37;
    }
    tagged_ptr _16(tagged_ptr * _38)
    {
        tagged_ptr _39 = mkClos(_18, 1, _38[0]);
        return _39;
    }
    tagged_ptr _29(tagged_ptr * _40)
    {
        tagged_ptr _41 = mkClos(_16, 0);
        tagged_ptr _42 = fixedPoint(_41);
        tagged_ptr _43 = mkZero();
        tagged_ptr _49 = inc(_43);
        tagged_ptr _50 = apply(_42, _49);
        tagged_ptr _51 = mkZero();
        tagged_ptr _56 = inc(_51);
        tagged_ptr _57 = apply(_50, _56);
        return _57;
    }
    int main()
    {
        call(_29);
    }

Which when run with preamble.c pasted on top it prints out 2. As you'd hope.