#esolang #compile-time #programs #macro #plugin #convert #fractran

nightly fractran_macros

A compiler plugin that converts Fractran code into Rust at compile time, letting your Fractran programs be optimised by LLVM to super-fast native code

5 releases

Uses old Rust 2015

0.1.5 Mar 21, 2015
0.1.4 Jan 26, 2015
0.1.3 Jan 4, 2015
0.1.1 Nov 20, 2014
0.1.0 Nov 14, 2014

#1044 in Programming languages

29 downloads per month

MIT/Apache

15KB
260 lines

fractran_macros

Build Status

A Rust macro for compiling FRACTRAN programs embedded in a Rust program into efficient, allocation-less, libcore-only1 code at compile time.

FRACTRAN is a very simple language; a program is an integer n along with a list of positive fractions, executed by finding the first fraction f for which nf is an integer, replace n by nf and repeating (execution halts when there is no such fraction). It turns out that this is Turing complete, and people have even written FRACTRAN interpreters in FRACTRAN! (See examples/fractran.rs.)

1That's right; you can now use FRACTRAN inside a kernel.

Usage

The fractran macro takes a series of comma-separated arithmetic expressions, representing the sequence of fractions. Supported operations:

  • *, / and parentheses for grouping; no risk of arithmetic overflow or loss of precision,
  • +; can overflow,
  • integer powers via ^; no overflow, but precedence is incorrect, so a^b * c is a^(b * c), rather than (a^b) * c as it should be. Use parentheses to ensure correctness.

The macro returns a constructor function fn(&[u32]) -> T, where &[u32] is the initial number (in the format described below), and T is a type implementing Iterator<()> and fractran_support::Fractran. Calling next will step the machine (i.e. finding the appropriate fraction as described above), returning None when the machine has halted.

The fractran_support::Fractran trait provides the state method (for viewing the current number, in exponent form) and the run method for executing the state machine until it halts.

Example

#![feature(phase)]

#[phase(plugin)] extern crate fractran_macros;

fn main() {
    // takes 2^a 3^b to 3^(a+b)
    let add = fractran!(3/2);
    println!("{}", add(&[12, 34]).run()); // [0, 46]

    // takes 2^a 3^b to 5^(ab)
    let mult = fractran!(455 / 33, 11/13, 1/11, 3/7, 11/2, 1/3);
    println!("{}", mult(&[12, 34]).run()); // [0, 0, 408, 0, ...]
}

Remember to ensure the Cargo.toml has the appropriate dependencies:

[dependencies.fractran_macros]
git = "https://github.com/huonw/fractran_macros"

Representation

Numbers are represented in terms of the (32-bit) exponents of their prime factors. The ith entry of [u32] values is the exponent of the ith prime (zero-indexed), for example 2 == [1], 3 == [0, 1], 9438 == 2 * 3 * 11^2 * 13 == [1, 1, 0, 0, 2, 1]. This representation allows the implementation to handle very large numbers, anything where the largest exponent of any prime is less than 232, so the smallest non-representable number is 2232 = 24294967296.

This also allows the internal implementation to be highly efficient with just (statically determined) array indexing and integer comparisons & additions; there is no possibility of out-of-bounds indexing (and thus no performance penalty from unwinding), nor is there any division or remainder operations. As an example, the example/prime.rs program uses Conway's prime enumeration FRACTRAN program to generate primes, it takes only 6 seconds to do 1 billion steps (reaching the lofty heights of 887).

Converting this representation to and from an actual number can be achieved with your favourite prime-number crate (e.g. zipping with the primes iterator from slow_primes).

Why?

Why not? Esolang-macros are fun, and so is the fundamental theorem of arithmetic.

Dependencies

~565KB
~11K SLoC