Lessons learned building a toy compiler
I’ve been working on Olifant; a compiler for simply typed lambda calculus for last few weeks, and this is an introduction to things that I learned along the way. Lambda Calculus is an extremely minimal formal system to study programming language theory and computation in general. Olifant compiles a slightly modified form of lambda Calculus into LLVM IR and then to machine code which can be run natively.
We tend to think of compilers as big black boxes which transform some high level language, let’s say C into a binary in one big step. I’d like to present it as a pipeline of languages and transformations, each a bit simpler and slightly lower level than the one before it; a complex compiler built out of a series of transformations in which each of them is a simple independent pure function.
The big picture
The language is called Calculus and it looks like this.
let id = λx:i.x
let k = 42
id k
λx is a function that takes one argument x. The :i indicates that x is
of type integer. Everything after the . is the body of the function and here
it is just the variable x. The value of the body is implicitly returned. id
k is a function application with a single argument k. Calculus is a fairly
simple language; a program is a series of function or variable declarations
(also called let bindings) and an expression in the end. The result of the
expression is returned from the program as the exit code.
The LLVM IR is a language low level enough to be close to metal but still expressive enough to describe a lot of high level features without too much verbosity. It is best thought of as a high level machine independent portable assembly. It is an excellent tool chain for compiler authors with great tooling, descriptive error messages, blazing fast performance and sophisticated optimizations built in. Targeting LLVM IR instead of x86 assembly turned out to be a very good idea in retrospect. See the language reference for a detailed description or follow along for simple examples.
Calculus can be compiled into LLVM IR with the olifant compiler
$ echo 'let id = λx:i.x; let k = 42; id k' | olifant
This is the equivalent LLVM IR
; ModuleID = 'calc'
@k = global i64 42
define i64 @id(i64 %x) {
entry:
ret i64 %x
}
define i64 @main(i64 %_) {
entry:
%0 = load i64, i64* @k
%1 = call i64 @id(i64 %0)
ret i64 %1
}
The llc compiler can compile the IR into an executable binary and optionally apply several optimizations with fine grained control and provide the machine assembly. However using clang is even simpler
$ clang sample.ll -o sample
$ file sample
sample: ELF 64-bit LSB executable, x86-64, version 1 (SYSV),
dynamically linked, interpreter /lib64/ld-linux-x86-64.so.2,
for GNU/Linux 2.6.32, not stripped
The Calculus is so small; it can’t even print the result to the terminal yet :)
The return value of a program can be fetched using bash special variables. $?
holds the exit code of the last executed program.
$ ./sample
$ echo $?
42
Parser
The quintessential first step in any compiler is parsing the source string into an Abstract Syntax Tree; which is just a convenient data structure to work with rather than raw strings. Olifant uses the venerable combinatorial parsing library Parsec and the paper Monadic Parsing in Haskell is a beautiful and simple introduction to the subject.
Consider a very simple language with just numbers and addition. AST for such a language can be described very elegantly with the following Haskell type:
data Expr = Number Int | Plus Expr Expr
A sample program like 6 + 4 can then be parsed into
Plus (Number 6) (Number 4)
The Calculus is so small, it’s type can be represented in just a few lines of Haskell. All Haskell code examples are for illustration purposes only and it is not strictly necessary to understand it to read further.
data Calculus =
Var Text
| Number Int
| Bool Bool
| App Calculus Calculus
| Lam Text Tipe Calculus
| Let Text Calculus
Calculus is partly typed. Function arguments are type annotated and the rest of
it can be inferred (more on this shortly). The parser is reasonably straight
forward and is only about 100 lines of very well documented code. It
is very lenient in general and can only report syntax errors, like let a =;
and won’t care about an undefined variable.
Parser parses the input expression let id = λx:i.x; let k = 42, id k into
[
Let "id" (Lam "x" TInt (Var "x"))
, Let "k" (Number 42)
, App (Var "id") (Var "k")
]
The 1 to 1 correspondence between the source language and the Calculus type is intentional.
Core
The second language we describe is called Core and is very different from Calculus. The BNF grammar for Calculus is beautifully recursive but Core clearly distinguishes what is allowed at the top level vs inside a nested expression. For example, functions can be defined only at the top level of Core, but they are allowed anywhere in Calculus. The limitations might seem arbitrary and unnecessarily crippling, but it makes subsequent analysis and code generation much simpler.
The Core is statically typed with the type of every expression known at compile time. TArrow is a function type.
data Tipe = TUnit | TInt | TBool | TArrow Tipe Tipe
An Expr is a nested Core expression.
data Expr
= Var Ref
| Lit Literal
| App Tipe Expr Expr
| Lam Tipe Ref Expr
A let binding is a top level expression of the form let <name> = <expr>
data Bind = Bind Ref Expr
A program is a list of bindings and a terminating expression
data Progn = Progn [Bind] Expr
Olifant can dump the core output with the -c flag and you can optionally
pretty print the output with hindent.
$ echo 'let id = λx:i.x; let k = 42; id k' | olifant -c | hindent
Progn
[ Bind
(Ref {rname = "id", raw = "id", rtipe = TArrow TInt TInt, scope = Global})
(Lam
(TArrow TInt TInt)
(Ref {rname = "x", raw ="x", rtipe = TInt, scope = Local})
(Var (Ref {rname = "x", raw ="x", tipe = TInt, scope = Local})))
, Bind (Ref {rname = "k", raw = "k", rtipe = TInt, scope = Global}) (Lit (LNumber 42))
]
(App
TInt
(Var (Ref {rname = "id", raw = "id", rtipe = TArrow TInt TInt, scope = Global}))
(Var (Ref {rname = "k", raw = "k", rtipe = TInt, scope = Global})))
References
Before we move on further, we need to talk about one of the most commonly used
types in Olifant, Ref used to represent a variable.
data Ref = Ref { name :: Text
, raw :: Text
, tipe :: Tipe
, scope :: Scope}
One of the interesting things I learned about compilers is that a symbol table is often unnecessary. Typically a symbol table is a mapping from a variable name to it’s properties like the value, type and the line number at it was defined.
Quoting AOSA § No Symbol Table,
In GHC we use symbol tables quite sparingly; mainly in the renamer and type checker. As far as possible, we use an alternative strategy: a variable is a data structure that contains all the information about itself. Indeed, a large amount of information is reachable by traversing the data structure of a variable: from a variable we can see its type, which contains type constructors, which contain their data constructors, which themselves contain types, and so on.
I find this approach fundamentally simpler and a natural way to think in a
purely functional programming language. Rather than represent a variable with
just a string and maintain a shared mapping from the name to all its attributes,
we use a rich type (Ref) to embed all the attributes in itself. Each pass of
the compiler can augment and annotate more information into the reference
accordingly. Large parts of the program become inherently stateless and simpler.
Cast
Casting is applying the minimum transformations on Calculus to get Core and is
the immediate step after parsing. Cast does basic structural validations and
ensures that the program is a series of declarations and an expression in the
end. All known types (only function arguments for now) are retained and the rest
of them are explicitly marked to be of unit type (think of Java’s Object).
Types of literal values are inferred here. Cast either returns a Core expression
or raises a syntax error.
Rename
Core to Core transformations in general preserve semantics but simplify them for
further transformations and code generation. The rename phase rewrites variable
names to avoid runtime ambiguity and shadowing and eliminates the need for
complex symbol tables. The program let x = 1; let id = λx.x uses the variable
x twice, but they are in no ways related. It is equivalent to let x = 1; let
id = λy.y but latter is simpler in a way that transformations after renamer can
be sure that variable references are globally unique and there is no need to
deal with name collisions.
The user defined name of the reference is preserved in a Ref as raw and a
new globally unique name is chosen (called name) is used from then onward.
It is an open question if the two approaches are the same and I haven’t figured out the answer yet.
Type inference
Type inference for a language with some type annotations and no polymorphic types is reasonably straightforward and easy, and I decided to let go of some expressive power (read generics) to keep things simple.
A polymorphic type or a type variable is to a type what a variable is to a
value. You describe the properties of an item from a set and it takes a concrete
value at runtime. For example, the identity function in Haskell has the type
signature id :: a -> a, which can be read as a function that takes a value of
polymorphic/generic type a and returns another value of type a. At runtime,
a gets a concrete value like Int or String, and polymorphic types let us
work with a set of types in a generic way. tldr; I don’t have polymorphic types
yet.
Complete type inference is possible for a language with polymorphic types even without any user annotations with a very well known type system called called Hindley-Milner with algorithm W but it is a bit tricky to get right. In the meanwhile, we can use a much simpler algorithm.
Instead of getting into the formal model, the algorithm can be explained using a
simple heuristic. The problem is to find the type of each expression in the
program without the user explicitly annotating everything. Think of the auto
keyword in C++. The idea is to discover that the language can introduce a new
variable in only 2 ways, function arguments and let bindings. As long as all
the parameters of the function are annotated, this information can be used in
the body of the function and need not be repeated. The type of a variable
introduced by let binding can be inferred from the right hand side of the
equation recursively, which is either a term constructed with known types or a
type error. Types of literals like numbers, booleans and strings are obvious
after parsing. In practice this approach worked really well for Olifant.
Code generation
The code generator accepts a well typed Core as input and generates LLVM IR. This module took me the longest to write, test and then rewrite several times.
I had to spend a lot of time learning the LLVM nuances and fight with the Haskell API before things started working. The Haskell bindings for LLVM provides a thin wrapper around the C++ API and it is nowhere close to being typesafe, so the usual Haskell magic of “if it compiles, it works” wasn’t applicable at all. I took so many failed approaches before I settled down with what I have right now.
The basic idea is to initialize a module, and to populate it with definitions as you traverse the AST. A module contains a list of global definitions and each function or global variable is translated into one. Global definitions are composed of blocks and a block constituents of a logical unit before a branch instruction like a conditional or function return.
The code is not hard to understand but uses some reasonably advanced Haskell machinery like monad transformers to stay clean and maintainable. The LLVM IR is printed to stdout and can be compiled to machine code with clang as explained in the beginning of the post.
Things that need to be talked about
There are a whole lot of things I had to omit from this post to save it from growing into a 200 page text book, but a few important ones need to be mentioned:
1. Lambda lifting
Lambda lifting is the technique used to lift inline function definitions to top level for easy code generation. This is necessary for the LLVM backend.
Further reading:
2. Single Static Assignment form
SSA is a property of an intermediate representation (IR), which requires that each variable is assigned exactly once, and every variable is defined before it is used. It sounds too simple and obvious, but the implications are amazing. This form makes a lot of optimizations like dead code elimination significantly simpler.
3. Left recursion in combinatorial parsers
I learned about this issue when my parser started going into an infinite loop for specific inputs. This blog post Left-recursion in Parsec is a decent explanation of the problem.
4. Learn LLVM with C
An excellent way to learn LLVM is to compile very simple C programs with optimizations turned off. The examples folder contains a sample C file and a Makefile with the right compiler flags.
Things that didn’t work
I went down way too many unsuccessful rabbit holes to list here, but here are a few
1. The one true perfect type safe AST probably doesn’t exist
I spent countless sleepless nights on my bed thinking about this problem and still didn’t achieve a satisfactory solution. Can you define your AST in a way that all semantic errors become type errors? This blog post The AST Typing Problem by Edward Z. Yang is a good introduction. I might write another post sometime later explaining a few common strategies.
2. Untyped lambda calculus is way harder
Untyped lambda calculus forces polymorphic types on you. You don’t want it when you are just starting out.
3. Using a pretty printer library to handle malformed AST is a bad idea
I was pretty stupid and got warned by Stephen Diehl. I learned my lesson.
Next steps
1. FFI
A foreign function interface to C will make the language a lot more useful and powerful; like printing to the terminal. Sigh!
Reading list
-
David Terei’s slides on the GHC compiler internals. Most of my ideas about tiny passes where shaped by Haskell.
-
The GHC user’s guide is one of the best places to learn Haskell
-
LLVM types can be very intimidating. I wrote a simple small list of types you might need.
You might find the discussions on Hacker News and Reddit useful.