Andrew Kelley

Introduction to the Zig Programming Language

2016 February 08

The past few months I took a break from working on Genesis Digital Audio Workstation to work, instead, on creating a new programming language.

I am nothing if not ambitious, and my goal is to create a new programming language that is more pragmatic than C. This is like to trying to be more evil than the devil himself.

So, in order, these are the priorities of Zig:

  1. Pragmatic: At the end of the day, all that really matters is whether the language helped you do what you were trying to do better than any other language.
  2. Optimal: The most natural way to write a program should result in top-of-the-line runtime performance, equivalent to or better than C. In places where performance is king, the optimal code should be clearly expressible.
  3. Safe: Optimality may be sitting in the driver's seat, but safety is sitting in the passenger's seat, wearing its seatbelt, and asking nicely for the other passengers to do the same.
  4. Readable: Zig prioritizes reading code over writing it. Avoid complicated syntax. Generally there should be a canonical way to do everything.

Table of Contents

  1. Table of Contents
  2. Design Decisions
    1. Widely Diverging Debug and Release Builds
    2. C ABI Compatibility
    3. Maybe Type Instead of Null Pointer
    4. The Error Type
    5. Alternate Standard Library
    6. Alternatives to the Preprocessor
  3. Milestone: Tetris Implemented in Zig
  4. Resources

Design Decisions

Widely Diverging Debug and Release Builds

Zig has the concept of a debug build vs a release build. Here is a comparison of priorities for debug mode vs release mode:

Debug Mode Release Mode
Time Spent Compiling Code must compile fast. Use all manner of caching, shared objects, multithreading, whatever must be done in order to produce a binary as soon as possible. Making a release build could take orders of magnitude longer than a debug build and that is acceptable.
Runtime Performance Could be order of magnitude slower than release build and that is acceptable. Optimal performance. Aggressive optimizations. Take the time needed to produce a highly efficient runtime efficiency. No compromises here.
Undefined Behavior What would be undefined behavior in a release build, is defined behavior in a debug build, and that is for the runtime to trap. That is, crash. This includes things like array bounds checking, integer overflow, reaching unreachable code. Not all undefined behavior can be caught, but a comfortably large amount can. Undefined behavior in release mode has unspecified consequences, and this lets the optimizer produce optimal code.

The build mode is available to the source code via the expression @compile_var("is_release").

Complete C ABI Compatibility

Part of being pragmatic is recognizing C's existing success. Interop with C is crucial. Zig embraces C like the mean older brother who you are a little afraid of but you still want to like you and be your friend.

In Zig, functions look like this:

fn do_something() {
    // ...

The compiler is free to inline this function, change its parameters, and otherwise do whatever it wants, since this is an internal function. However if you decide to export it:

export fn do_something() {
    // ...

Now this function has the C ABI, and the name shows up in the symbol table verbatim. Likewise, you can declare an external function prototype:

extern fn puts(s: &const u8) -> c_int;

In Zig, like in C, you typically do not create a "wrapper" or "bindings" to a library, you just use it. But if you had to type out or generate all the extern function prototypes, this would be a binding. That is why Zig has the ability to parse .h files:

c_import {

This exposes all the symbols in stdio.h - including the macros - to the zig program, and then you can call puts or printf just like you would in C.

One of Zig's use cases is slowly transitioning a large C project to Zig. Zig can produce simple .o files for linking against other .o files, and it can also generate .h files based on what you export. So you could write part of your application in C and part in Zig, link all the .o files together and everything plays nicely with each other.

Maybe Type Instead of Null Pointer

One area that Zig provides safety without compromising efficiency or readability is with the maybe type.

The question mark symbolizes the maybe type. You can convert a type to a maybe type by putting a question mark in front of it, like this:

// normal integer
const normal_int: i32 = 1234;

// maybe integer
const maybe_int: ?i32 = 5678;

Now the variable maybe_int could be an i32, or null.

Instead of integers, let's talk about pointers. Null references are the source of many runtime exceptions, and even stand accused of being the worst mistake of computer science.

Zig does not have them.

Instead, you can use a maybe pointer. This secretly compiles down to a normal pointer, since we know we can use 0 as the null value for the maybe type. But the compiler can check your work and make sure you don't assign null to something that can't be null.

Typically the downside of not having null is that it makes the code more verbose to write. But, let's compare some equivalent C code and Zig code.

Task: call malloc, if the result is null, return null.

C code

// malloc prototype included for reference
void *malloc(size_t size);

struct Foo *do_a_thing(void) {
    char *ptr = malloc(1234);
    if (!ptr) return NULL;
    // ...

Zig code

// malloc prototype included for reference
extern fn malloc(size: size_t) -> ?&u8;

fn do_a_thing() -> ?&Foo {
    const ptr = malloc(1234) ?? return null;
    // ...

Here, Zig is at least as convenient, if not more, than C. And, the type of "ptr" is &u8 not ?&u8. The ?? operator unwrapped the maybe type and therefore ptr is guaranteed to be non-null everywhere it is used in the function.

The other form of checking against NULL you might see looks like this:

void do_a_thing(struct Foo *foo) {
    // do some stuff

    if (foo) {

    // do some stuff

In Zig you can accomplish the same thing:

fn do_a_thing(maybe_foo: ?&Foo) {
    // do some stuff

    if (const foo ?= maybe_foo) {

    // do some stuff

Once again, the notable thing here is that inside the if block, foo is no longer a maybe pointer, it is a pointer, which cannot be NULL.

One benefit to this is that functions which take pointers as arguments can be annotated with the "nonnull" attribute - __attribute__((nonnull)) in GCC. The optimizer can sometimes make better decisions knowing that pointer arguments cannot be null.

The Error Type

Probably the most distinguishing feature of Zig is its exception handling strategy.

One of the top level declarations you can do is the error value declaration:

error FileNotFound;
error OutOfMemory;
error UnexpectedToken;

These error values are assigned an unsigned integer value greater than 0 at compile time. You are allowed to declare the same error value more than once, and if you do, it gets the same error value.

You can refer to these error values with the error namespace such as error.FileNotFound. The special value error.Ok has value 0.

Each error value across the entire compilation unit gets a unique integer, and this determines the size of the pure error type.

The pure error type is one of the error values, and in the same way that pointers cannot be null, a pure error cannot be error.Ok.

const pure_error = error.FileNotFound;

Most of the time you will not find yourself using a pure error type. Instead, likely you will be using the error union type. This is when you take a normal type, and prefix it with the % operator.

Here is parse_u64 from the standard library:

pub error InvalidChar;
pub error Overflow;

pub fn parse_u64(buf: []u8, radix: u8) -> %u64 {
    var x : u64 = 0;

    for (buf) |c| {
        const digit = char_to_digit(c);

        if (digit >= radix) {
            return error.InvalidChar;

        // x *= radix
        if (@mul_with_overflow(u64, x, radix, &x)) {
            return error.Overflow;

        // x += digit
        if (@add_with_overflow(u64, x, digit, &x)) {
            return error.Overflow;

    return x;

Notice the return type is %u64. This means that the function either returns an unsigned 64 bit integer, or an error.

Within the function definition, you can see some return statements that return a pure error, and at the bottom a return statement that returns a u64. Both types implicitly cast to %u64.

What it looks like to use this function varies depending on what you're trying to do. One of the following:

If you want to provide a default value, you can use the %% binary operator:

fn do_a_thing(str: []u8) {
    const number = parse_u64(str, 10) %% 13;
    // ...

In this code, number will be equal to the successfully parsed string, or a default value of 13. The type of the right hand side of the binary %% operator must match the unwrapped error union type, or be unreachable.

Let's say you wanted to return the error if you got one, otherwise continue with the function logic:

fn do_a_thing(str: []u8) -> %void {
    const number = parse_u64(str, 10) %% |err| return err;
    // ...

There is a shortcut for this. The %return expression is equivalent:

fn do_a_thing(str: []u8) -> %void {
    const number = %return parse_u64(str, 10);
    // ...

%return evaluates an error union expression, and then returns the same error if it was an error. Otherwise, returns the unwrapped value.

Maybe you know with complete certainty that an expression will never be an error. In this case you can do this:

const number = parse_u64("1234", 10) %% unreachable{};

Here we know for sure that "1234" will parse successfully. So we put the unreachable{} value on the right hand side. unreachable{} generates a panic in debug mode and undefined behavior in release mode. So, while we're debugging the application, if there was a surprise error here, the application would crash appropriately.

Again there is a syntactic shortcut for this:

const number = %%parse_u64("1234", 10);

The %% prefix operator is equivalent. It unwraps an error union type, and panics in debug mode if the value was an error.

Finally, you may want to take a different action for every situation. For that, we have the switch expression:

fn do_a_thing(str: []u8) {
    switch (parse_u64(str, 10)) {
        error.Ok => |number| do_something_with_number(number),
        error.Overflow => {
            // handle overflow...
        // we promise that InvalidChar won't happen (or crash in debug mode if it does)
        error.InvalidChar => unreachable{},

The other component to error handling is defer statements. In addition to an unconditional defer, Zig has %defer, which evaluates the deferred expression on block exit path if and only if the function returned with an error from the block.


fn create_foo(param: i32) -> %Foo {
    const foo = %return try_to_allocate_foo();
    // now we have allocated foo. we need to free it if the function fails.
    // but we want to return it if the function succeeds.
    %defer deallocate_foo(foo);

    const tmp_buf = allocate_tmp_buffer() ?? return error.OutOfMemory;
    // tmp_buf is truly a temporary resource, and we for sure want to clean it up
    // before this block leaves scope
    defer deallocate_tmp_buffer(tmp_buf);

    if (param > 1337) return error.InvalidParam;

    // here the %defer will not run since we're returning success from the function.
    // but the defer will run!
    return foo;

The neat thing about this is that you get robust error handling without the verbosity and cognitive overhead of trying to make sure every exit path is covered. The deallocation code is always directly following the allocation code.

A couple of other tidbits about error handling:

Alternate Standard Library

Part of the Zig project is providing an alternative to libc.

libc has a lot of useful stuff in it, but it also has cruft. Since we're starting fresh here, we can create a new API without some of the mistakes of the 70s still haunting us, and with our 20-20 hindsight.

Further, calling dynamically linked functions is slow. Zig's philosophy is that compiling against the standard library in source form is worth it. In C this would be called Link Time Optimization - where you generate Intermediate Representation instead of machine code and then do another compile step at link time. In Zig, we skip the middle man, and create a single compilation unit with everything in it, then run the optimizations.

So, you can choose to link against libc and take advantage of it, or you can choose to ignore it and use the Zig standard library instead. Note, however, that virtually every library you depend on probably also depends on libc, which drags libc as a dependency into your project. Using libc is still a first class use case for Zig.

Alternatives to the Preprocessor

The C preprocessor is extremely powerful. Maybe a little too powerful.

The problem with the preprocessor is that it turns one language into two languages that don't know about each other.

Here are some examples of where the preprocessor messes things up:

Regardless of the flaws, C programmers find ourselves using the preprocessor because it provides necessary features, such as conditional compilation, a constant that can be used for array sizes, and generics.

Zig plans to provide better alternatives to solve these problems. For example, the constant expression evaluator of Zig allows you to do this:

const array_len = 10 * 2 + 1;
struct Foo {
    array: [array_len]i32,

This is not an amazing concept, but it eliminates one use case for #define.

Next, conditional compilation. In Zig, all the compilation variables are available via the builtin function @compile_var.

When you use it, it evaluates to a constant expression which knows its value is dependant on a compilation variable. Then you can just write normal code:

fn do_something() {
    if (@compile_var("is_release")) {
        // do the release behavior
    } else {
        // do the debug behavior

This is guaranteed to leave out the if statement when the code is generated. However, the semantic analyzer still runs on all the code.

One use case for conditional compilation is demonstrated in libsoundio:

static const enum SoundIoBackend available_backends[] = {

Here, we want a statically sized array to have different contents depending on whether we have certain libraries present.

In Zig, it would look something like this:

const available_backends =
    (if (@compile_var("have_jack"))
    (if (@compile_var("have_pulse_audio"))
    (if (@compile_var("have_alsa"))
    (if (@compile_var("have_core_audio"))
    (if (@compile_var("have_wasapi"))

Here we take advantage of the compile-time array concatenation operator, ++. It's a bit more verbose than the C equivalent, but the important thing is that it's one language, not two.

Finally, generics. Zig does not have generics yet, but has plans for type-generic functions and data structures that should be the nail in the coffin for the preprocessor. Macros are on the table but no current plan for how to do them.

Milestone: Tetris Implemented in Zig

This past week I achieved a fun milestone: a fully playable Tetris clone implemented in Zig, with the help of libc, GLFW, and libpng.

If you're using Linux on the x86_64 architecture, which is currently the only supported target, you could build Zig and then build this Tetris game.

Otherwise, here's a video of me demoing it:


If you are interested in the language, feel free to participate.