Brainfuck is an interesting programming language, to say the least. As the most prominent Esoteric programming language out there, it provides a truly cursed programming experience. It is quite minimalistic as well, featuring only 8 distinct instructions and a programming model that can be best described as a glorified Turing machine, making writing a Brainfuck compiler an interesting programming exercise. However, we might as well go a step further and develop an intermediate representation for it, because, well, why not?1

The Brainfuck programming model: a refresher Link to heading

In order to execute a Brainfuck program, four crucial elements are needed:

  1. The program source code (the only element that is immutable)
  2. The program counter
  3. The memory (an infinite2 array of byte-sized cells)
  4. The memory pointer

Here is a Rust example of what a Brainfuck “VM” might look like:

/// The Brainfuck Virtual Machine holds all the data
/// necessary to compile and run a Brainfuck program
struct BFVM {
    memory: [u8; 300000],
    mem_pointer: usize,

    code: Vec<OpCode>,
    code_pointer: usize,
}

The Brainfuck instruction set Link to heading

Brainfuck puts the Reduced in RISC, featuring a grand total of 8 instructions:

  • > - Moves the memory pointer one cell to the right
  • < - Moves the memory pointer one cell to the left
  • + - Adds 1 to the value of the cell
  • - - Subtracts 1 to the value of the cell
  • . - Outputs the value from the cell to the command line
  • , - Reads a character from the command line
  • [ - jumps to the matching ], if the value of the cell is 0
  • ] - jumps to the matching [, if the value of the cell is not 0

If you are looking to better understand some of Brainfuck’s nuances, I very much recommend this informal specification, from which the instruction set subchapter is mostly based on.

Potential for improvement Link to heading

For what Brainfuck makes up in simplicity, it loses out in expressiveness and efficiency (or a lack thereof). Please take a look at the code excerpt below:

+++++++++++++[->++>>>+++++>++>+<<<<<<]>>>>>++++++>--->>>>>>>>>>+++++++++++++++[[
>>[-<<<<<<+>>>>>>]<<<<<<[->>>>>>+<<+<<<+<]>>>>>>>>]<<<<<<<<<[<<<<<<<<<]>>>>>>>>>
[>>>>>>>>[-<<<<<<<+>>>>>>>]<<<<<<<[->>>>>>>+<<+<<<+<<]>>>>>>>>]<<<<<<<<<[<<<<<<<
<<]>>>>>>>[-<<<<<<<+>>>>>>>]<<<<<<<[->>>>>>>+<<+<<<<<]>>>>>>>>>+++++++++++++++[[
>>>>>>>>>]+>[-]>[-]>[-]>[-]>[-]>[-]>[-]>[-]>[-]<<<<<<<<<[<<<<<<<<<]>>>>>>>>>-]+[

When I look at Brainfuck code, the thing that strikes me most is how repetitive it looks. For example, the first line starts with 13 + instructions in a row. I could just as easily find examples for the other three memory-manipulating instructions (- < >).

Another pattern that comes up often is this one: [-]. This loop decrements the value of the cell until it reaches 0.

Lastly, finding the matching ] for the respective [ (or vice-versa) can be very time-consuming, depending on how much of your program is enclosed by this bracket pair.

An improved intermediate representation Link to heading

Taking these inefficiencies into account, we can develop a higher-level and more performant abstraction:

/// Brainfuck source code gets compiled
/// to an intermediate representation made up of `OpCode`s
enum OpCode {
    /// Increments the value in the cell by x (can overflow)
    Increment(u8),
    /// Decrements the value in the cell by x (can underflow)
    Decrement(u8),
    /// Moves the pointer x values to the left
    MoveLeft(usize),
    /// Moves the pointer x values to the right
    MoveRight(usize),
    /// Sets the value in cell to 0
    Zero,
    /// Reads a value into the cell
    Read,
    /// Prints the value from the cell as an ASCII character
    Write,
    /// Jump to destination if:
    /// - The value from the cell is 0 and the direction is [Direction::Forward]
    /// - The value from the cell is not 0 and the direction is [Direction::Backward]
    Jump {
        destination: usize,
        direction: Direction,
    },
}

#[derive(Eq, PartialEq)]
enum Direction {
    Forward,
    Backward,
}

As you can see, not only did we mitigated the repetitiveness of the memory-manipulating instructions but we also turned the [ ] jump instructions into a O(1) operation by precomputing the destination instruction index. An instruction that zeroes out the cell was also introduced, in order to replace the [-] pattern.

Compiling Brainfuck code Link to heading

Now that we designed our IR, we need to write a function that compiles the source code. We start by dealing with the repetitive memory-manipulating instructions (+ - < >):

impl BFVM {
    fn new(code: &str) -> Self {
        ...
    }

    fn compile(code: &Vec<char>) -> Vec<OpCode> {
        //! Compiles the `code` into IR i.e. a vector of [OpCode]s
        let mut res: Vec<OpCode> = Vec::new();
        let mut index = 0;
        let mut jumps: Vec<usize> = Vec::new();
    
        while index < code.len() {
            match code[index] {
                '<' | '>' | '+' | '-' => {
                    // These 4 operators can be be treated the same way:
                    // For example, if a sequence of "<<<<<" appears,
                    // they get condensed to a MoveLeft(5) opcode
                    let hit = code[index];
                    let mut len = 1;
                    while index + len < code.len() && code[index + len] == hit {
                        len += 1;
                    }
    
                    let opcode = match code[index] {
                        '<' => OpCode::MoveLeft(len),
                        '>' => OpCode::MoveRight(len),
                        '+' => OpCode::Increment(len as u8),
                        '-' => OpCode::Decrement(len as u8),
                        _ => panic!(),
                    };
    
                    res.push(opcode);
                    index += len;
                }

The Read and Write opcodes are quite straightforward to compile:

                '.' => {
                    res.push(OpCode::Write);
                    index += 1;
                }
                ',' => {
                    res.push(OpCode::Read);
                    index += 1;
                }

Whenever a [ instruction is found, we are dealing with one of two cases:

  • If we’re dealing with the [-], emit a Zero opcode
  • Otherwise, it’s just a Jump opcode. We keep track of this forward jump by pushing its index to the jumps stack, so that when we find the corresponding ] we can fill in the destination field.
                '[' => {
                    if index + 2 < code.len() && code[index + 1] == '-' && code[index + 2] == ']' {
                        // If this pattern "[-]" appears, it means it is just zeroing out a value
                        res.push(OpCode::Zero);
                        index += 3;
                    } else {
                        res.push(OpCode::Jump {
                            direction: Direction::Forward,
                            destination: 0,
                        });
                        index += 1;
    
                        jumps.push(res.len() - 1); // Updating the jumps stack so that
                                                   // we keep track of the index of this '['
                    }
                }

When we deal with a ] instruction, we need to make sure we have both jumps pointing at each other:

                ']' => {
                    let dest = jumps.pop().unwrap();
    
                    res[dest] = OpCode::Jump {
                        direction: Direction::Forward,
                        destination: res.len(),
                    };
    
                    res.push(OpCode::Jump {
                        direction: Direction::Backward,
                        destination: dest,
                    });
                    index += 1;
                }
                _ => index += 1, // A comment char, moving on
            }
        }
        res
    }
}

Running our intermediate representation Link to heading

Now that we compiled our source code, we need to run it. Luckily for us, Rust’s match statement excels at dealing with the various values of a enum:

impl BFVM {
    ...
    fn run(&mut self) {
        //! Runs the VM
        while self.code_pointer < self.code.len() {
            match &self.code[self.code_pointer] {
                // Memory values can and do overflow and underflow,
                // hence the use of overflowing_add and overflowing_sub
                OpCode::Increment(i) => {
                    self.memory[self.mem_pointer] =
                        self.memory[self.mem_pointer].overflowing_add(*i).0
                }
                OpCode::Decrement(i) => {
                    self.memory[self.mem_pointer] =
                        self.memory[self.mem_pointer].overflowing_sub(*i).0
                }
                OpCode::MoveLeft(i) => self.mem_pointer -= i,
                OpCode::MoveRight(i) => self.mem_pointer += i,
                OpCode::Zero => self.memory[self.mem_pointer] = 0,
                OpCode::Read => self.read_input(),
                OpCode::Write => print!("{}", self.memory[self.mem_pointer] as char),
                _ => (),
            }

The code pointer manipulation logic is handled separately to ensure that it is done correctly:

            if let OpCode::Jump {destination, direction} = &self.code[self.code_pointer] {
                let zero = self.memory[self.mem_pointer] == 0;
                if (direction == &Direction::Forward && zero)
                    || (direction == &Direction::Backward && !zero)
                {
                    self.code_pointer = *destination;
                } else {
                    self.code_pointer += 1;
                }
            } else {
                self.code_pointer += 1;
            }
        }
    }
    ...
}

Final remarks and a Mandelbrot set Link to heading

Sometimes, the most useless projects are the ones that are the most fun to work on. Well, I certainly enjoyed working on this compiler and I hope you enjoyed reading about it. Finally, you can find here the source code used for this blog post. I’ll leave you with an ASCII representation of the Mandelbrot set, generated by the compiler.

ASCII representation of the Mandelbrot set


  1. I stand in the shoulders of giants that have done a much better job than I have of optimizing Brainfuck. ↩︎

  2. 300000 cells is infinite enough. ↩︎