My First LLVM Compiler

A walkthrough of writing a basic compiler with LLVM. No prior experience assumed.

In a short space of time, I was able to go from zero C++ knowledge, and no experience with LLVM, to a fully-fledged compiler. It’s a lot of fun, let me show you how!

Our compiler will accept programs written in BF. This is a classic toy language for compilers, and there is even a BF compiler in LLVM’s examples directory! In this post, I’ll explain the process of writing it.

A first LLVM program

The LLVM toolchain is built around programs written in LLVM IR. First, we’ll write a basic LLVM IR program that just exits.

Since LLVM tools already use LLVM IR, we can simply use the clang compiler to show us what a basic program looks like. Here’s a simple C program:

int main() {
    return 42;

We run this through Clang:

# -O3 ensures we discard any unnecessary instructions.
$ clang -S -emit-llvm -O3 forty_two.c

This creates a file forty_two.ll which looks like this:

define i32 @main() {
  ret i32 42

Our first LLVM program! We can use lli to run it:

$ lli forty_two.ll 
$ echo $?

Writing a skeleton

We’ll compile BF commands to sequences of LLVM IR instructions. However, those instruction will need to be inside a main function so LLVM knows our entry point. We will also need to allocate and initialise our memory cells and cell index.

Again, we can simply write the equivalent C and look at the LLVM IR instructions Clang generates. Here’s the skeleton we’ll use:

declare i8* @calloc(i32, i32)
declare void @free(i8*)

define i32 @main() {
  ; Allocate 30,000 cells on the heap.
  %cells = call i8* @calloc(i32 30000, i32 1)

  ; Allocate a stack variable to track the cell index.
  %cell_index_ptr = alloca i32
  ; Initialise it to zero.
  store i32 0, i32* %cell_index_ptr

  ; Our BF code will go here!

  ; Free the memory for the cells.
  call void @free(i8* %cells)
  ret i32 0

Hand-compiling >

> is the easiest BF command to compile. Let’s fire up a text editor and write out the LLVM IR equivalent.

If your BF is rusty, > simply increments the cell index.

%cell_index = load i32* %cell_index_ptr
%new_cell_index = add i32 1, %cell_index
store i32 %new_cell_index, i32* %cell_index_ptr

To check this code is correct, we want to run it. We can simply wrap it in our skeleton and run with it with lli to see what happens. Implementing < is now straightforwards, and we write a test program for < too.

Hand-compiling +

BF’s + command increments the current cell. This requires us to dereference the current cell, calculate new value, then store it. In C, this would look like:

char *cell_ptr = cells + cell_index;
char current_value = *cell_ptr;
char new_value = current_value + 1;
*cell_ptr = new_value;

LLVM provides the getelementptr instruction to calculate the pointer. The translation then looks like this:

%cell_index = load i32* %cell_index_ptr
%cell_ptr = getelementptr i8* %cells, i32 %cell_index

%current_value = load i8* %cell_ptr
%new_value = add i8 %current_value, 1
store i8 %new_value, i8* %cell_ptr

Again, we test this by wrapping it in our skeleton, and do the same for -.


BF has two I/O commands: , reads from stdin into a cell, and . writes from a cell onto stdout. We need to call C functions for this: putchar and getchar.

We need to declare these functions, as we did with malloc earlier:

declare i32 @putchar(i32)
declare i32 @getchar()

To implement , we call getchar, truncate it to a char, and write it to the current cell.

%cell_index = load i32* %cell_index_ptr
%cell_ptr = getelementptr i8* %cells, i32 %cell_index

%input_int = call i32 @getchar()
%input_byte = trunc i32 %input_int to i8
store i8 %input_byte, i8* %cell_ptr

. is the reverse: we read the cell, sign extend it, then call putchar.

%cell_index = load i32* %cell_index_ptr
%cell_ptr = getelementptr i8* %cells, i32 %cell_index

%current_cell = load i8* %cell_ptr
%current_cell_word = sext i8 %current_cell to i32
call i32 @putchar(i32 %current_cell_word)


LLVM IR instructions are organised into basic blocks; sequences of instructions that don’t contain jumps. At the end of each basic block, you must either jump to another basic block or return.

To compile [ x ] y, we need to inspect the current cell, then either jump to x, which is our loop body, or to y. At the end of x, we need to jump back to the start.

  %cell_index = load i32* %cell_index_ptr
  %cell_ptr = getelementptr i8* %cells, i32 %cell_index
  %cell_value = load i8* %cell_ptr
  %is_zero = icmp eq i8 %cell_value, 0
  br i1 %is_zero, label %loop_after, label %loop_body

  ; x
  br label %loop_header

  ; y

Note that this is recursive, x may also contain loops. There’s a sample loop program here.

Putting it all together

We’ve done the hard bit! All that’s left is to use the LLVM API generate these instructions. Each IR instruction has a corresponding C++ object that you can instantiate and add to your basic block.

LLVM’s API also has the convenient concept of an IRBuilder. The IRBuilder class provides a create method for every IR instruction, making IR generation easy.

Here’s the C++ for generating LLVM IR for >. The excellent LLVM tutorial includes instructions for compiling a basic C++ program that uses LLVM APIs.

BasicBlock *BB;

// Instantiate an IRBuilder that appends to our
// current basic block.
IRBuilder<> Builder(getGlobalContext());

// We want to increment by 1, but since cell_index is
// 32-bit, our constant must be 32-bit too.
Value *IncrementAmount =
    ConstantInt::get(getGlobalContext(), APInt(32, 1));

// Emit the load, add and store instructions.
Value *CellIndex = Builder.CreateLoad(CellIndexPtr, "cell_index");
Value *NewCellIndex =
    Builder.CreateAdd(CellIndex, IncrementAmount, "new_cell_index");
Builder.CreateStore(NewCellIndex, CellIndexPtr);

Compiling the other BF commands is a simple translation of our hand-compiled snippets. You can view the full working compiler here.

Emitting machine code

Finally, our compiler is only emitting LLVM IR. A proper compiler emits machine code. We can use llc to convert to an object file, then link it to produce an executable.

$ ./compiler
$ llc -filetype=obj hello_world.ll
$ gcc hello_world.o -o hello_world
$ ./hello_world
Hello World!

And that’s all there is to it!

Addendum: Optimising

There’s lots that can be done to produce faster executable from BF programs. However, LLVM is already smart enough to compile simple loop-free BF programs to optimal LLVM IR!

$ cat 
+++++ +++++
+++++ +++++
+++++ +++++
+++ ASCII 33 is '!'
. Write ! to stdout
$ ./compiler 
$ opt -S -O3 exclamation.ll -o optimised_exclamation.ll

This produces:

define i32 @main() {
  %0 = tail call i32 @putchar(i32 33)
  ret i32 0

Recent Posts

Announcing Remacs: Porting Emacs to Rust

Introspecting Glue Code

Searching A Million Lines Of Lisp

Rustdoc Meets The Self-Documenting Editor

Example Driven Development