A Brainfuck to binary compiler using LLVM, written in OCaml

Taddeus Kroes ddb09a9dd7 Extended optimization example il y a 11 ans
.gitignore 541164a238 First working version of compiler -> llvm il y a 11 ans
Makefile 2eafc343a2 Switched to LLVM 3.4 il y a 11 ans
README.md ddb09a9dd7 Extended optimization example il y a 11 ans
bf.ml 50b028d2a0 Code + makefile cleanup il y a 11 ans
hello.b 7b66f09168 Removed undocumented Hello World program and added rot13 program il y a 11 ans
rot13.b 7b66f09168 Removed undocumented Hello World program and added rot13 program il y a 11 ans
run.sh 2eafc343a2 Switched to LLVM 3.4 il y a 11 ans
text.py d1d5c13d26 Added Python script to generate naive text programs il y a 11 ans

README.md

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.

Building and usage

Building the bf compiler (Debian dependencies):

$ sudo apt-get install ocaml llvm-3.4 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 genrates 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: .
    %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)
    %1 = tail call i32 @putchar(i8 101)
    %2 = tail call i32 @putchar(i8 108)
    %3 = tail call i32 @putchar(i8 108)
    %4 = tail call i32 @putchar(i8 111)
    %5 = tail call i32 @putchar(i8 32)
    %6 = tail call i32 @putchar(i8 87)
    %7 = tail call i32 @putchar(i8 111)
    %8 = tail call i32 @putchar(i8 114)
    %9 = tail call i32 @putchar(i8 108)
    %10 = tail call i32 @putchar(i8 100)
    %11 = tail call i32 @putchar(i8 33)
    %12 = tail call i32 @putchar(i8 10)
    tail call void @exit(i32 0)
    ret void
}