brainfucker

About

This is a minimal compiler for the Brainfuck language, written for the purpose of practicing OCaml LLVM bindings and seeing how well LLVM optimizes arrays. Brainfuck commands are transformed to LLVM IR, which is generated in such a way that it is easy to optimize for LLVM's opt utility.

The (only) source file bf.ml pretty much explains itself. hello.b and rot13.b can be used for quick testing as demonstrated below.

Building and usage

Building the bf compiler (Debian dependencies, replace 3.4 with the version of the llvm package):

$ sudo apt-get install ocaml llvm libllvm-3.4-ocaml-dev
$ make

Building a Brainfuck program (plain and optimized):

$ make hello hello-opt  # compile file "hello.b" to binaries
$ ./hello
Hello World!
$ ./hello-opt
Hello World!

Quick compilation and running (deletes temporary binary after running):

$ echo ++++++++++++. | ./run.sh
$ ./run.sh < hello.b
Hello World!

Examining generated LLVM:

$ echo ++++++++++++. | ./bf
...
$ echo ++++++++++++. | ./bf | opt -O3 -S
...

Optimization example

The text.py utility generates single-cell Brainfuck code for a given text:

$ ./text.py Hello World!
++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++.
+++++++++++++++++++++++++++++.
+++++++.
.
+++.
-------------------------------------------------------------------------------.
+++++++++++++++++++++++++++++++++++++++++++++++++++++++.
++++++++++++++++++++++++.
+++.
------.
--------.
-------------------------------------------------------------------.
-----------------------.

The compiler generates very verbose code:

$ ./text.py Hello World! | ./bf
...

define void @_start() {
entry:
    ; initialization
    %mem = alloca [30000 x i8]
    %idx = alloca i32
    %0 = bitcast [30000 x i8]* %mem to i8*
    call void @llvm.memset.p0i8.i32(i8* %0, i8 0, i32 30000, i32 0, i1 false)
    store i32 0, i32* %idx

    ; command: +
    %1 = load i32* %idx
    %2 = getelementptr inbounds [30000 x i8]* %mem, i32 0, i32 %1
    %3 = load i8* %2
    %4 = add i8 %3, 1
    %5 = load i32* %idx
    %6 = getelementptr inbounds [30000 x i8]* %mem, i32 0, i32 %5
    store i8 %4, i8* %6

    ; command: +
    %7 = load i32* %idx
    %8 = getelementptr inbounds [30000 x i8]* %mem, i32 0, i32 %7
    %9 = load i8* %8
    %10 = add i8 %9, 1
    %11 = load i32* %idx
    %12 = getelementptr inbounds [30000 x i8]* %mem, i32 0, i32 %11
    store i8 %10, i8* %12

    ...

    ; command: . (outputs 'H' after 72 times a '+' command)
    %433 = load i32* %idx
    %434 = getelementptr inbounds [30000 x i8]* %mem, i32 0, i32 %433
    %435 = load i8* %434
    %436 = call i32 @putchar(i8 %435)

    ...

    call void @exit(i32 0)
    ret void
}

The LLVM optimization engine is able to completely optimize away array accesses using constant propagation/folding. In the absence of loops, this effectively evaluates the whole program at compile time:

$ ./text.py Hello World! | ./bf | opt -O3 -S
...
define void @_start() {
entry:
    %0 = tail call i32 @putchar(i8 72)    ; H
    %1 = tail call i32 @putchar(i8 101)   ; e
    %2 = tail call i32 @putchar(i8 108)   ; l
    %3 = tail call i32 @putchar(i8 108)   ; l
    %4 = tail call i32 @putchar(i8 111)   ; o
    %5 = tail call i32 @putchar(i8 32)    ;
    %6 = tail call i32 @putchar(i8 87)    ; W
    %7 = tail call i32 @putchar(i8 111)   ; o
    %8 = tail call i32 @putchar(i8 114)   ; r
    %9 = tail call i32 @putchar(i8 108)   ; l
    %10 = tail call i32 @putchar(i8 100)  ; d
    %11 = tail call i32 @putchar(i8 33)   ; !
    %12 = tail call i32 @putchar(i8 10)   ; \n
    tail call void @exit(i32 0)
    ret void
}