Caleb Schoepp

Adding A List Data Type To Lox

Lists are an important data type for any kind of serious programming. The following is a detailed walkthrough of how to add lists to the interpreted programming language Lox. We’ll go over the entire design including things like bytecode representations and syntax grammars. You can also expect a lot of code snippets1 showing the actual implementation details.

Lox is a programming language from the fantastic book Crafting Interpreters by Bob Nystrom. If you haven’t read the book yet, I highly recommend it. This post assumes a level of familiarity with Lox and it’s C implementation, Clox2. But, even if you haven’t read the book, it should still be a valuable lesson on programming language design.

My hope is that this post will help give you the confidence to start hacking on programming languages yourself. When I finished Crafting Interpreters I certainly didn’t feel qualified to keep working on the language. I missed the comfort of Robert’s witty writing, delightful drawings, and copious code-snippets. I’m here to show you that if I can add something like lists to Lox then so can you.

Overview

When the Clox interpreter runs it starts by lexing source code into a stream of tokens. Next, it simultaneously parses the stream of tokens and compiles it into bytecode3. Finally, the VM executes the bytecode.

When I’m thinking about adding a new feature to a language, I tend to work roughly in the opposite direction that the interpreter does. My process looks something like this:

  1. Think about what the desired outcome of the feature will be and what the corresponding source code would look like. Or in other words, I outline the semantics of the feature.

  2. Design and implement the feature’s runtime. The Clox interpreter is really just a C program. Therefore, we must represent the feature’s semantics in some performant C code.

  3. Determine the necessary opcodes for the feature. Then in the VM, tie the decoding of these opcodes to the runtime.

  4. Formalize the grammar of the new feature’s syntax. With that in hand, write the scanning and parsing to compile bytecode for the feature.

I generally find working in reverse easier. When I start with something like lexing I have a hard time understanding how a couple of new tokens will help me get where I want to go. Flipping the script lets me pretend everything already works and then systematically fill it in. The one disadvantage is that starting at the end often makes testing harder until you have finished the entire feature.

With this in mind, let’s dive in!

Semantics and Source Code

Semantics is just a fancy way of saying what we want something to do. We should start with the basics. How do we want the user of our language to define a list literal?

// In an assignment
var foo = ["a", "b", "c"];

// An empty list
var bar = [];

// Directly in an expression
print[(1, 2, 3)]; // Expect: [1, 2, 3]

Commonly you’ll see the items of large lists defined across multiple lines. In many languages, it is idiomatic to include a trailing comma. Let’s support that.

var foo = [1, 2, ...25];

If all we could do with lists is define them at compile time then they would be pretty useless. We want to be able to access items in the list. We’d also like to be able to store new values into the list.

var foo = [1, 2];

print foo[0]; // Expect: 1

foo[0] = 2;

print foo[0]; // Expect: 2

Our list type will be passed by reference. This means that functions will receive references to list arguments instead of copies. As a result, a function can modify a list without returning it.

fun modifyList(list) {
    list[0] = 0;
}

var foo = [1, 2];
modifyList(foo);
print foo; // Expect: [0, 2]

Up until this point, we’ve defined the semantics of something more like an array. It has a fixed-length set at compile-time, and it contains only a single type of value. To make this list a bit more “listy”, let’s start by saying it can store a mix of any type.

var foo = [1, "b", false];

As a convenience, we also want to be able to use expressions when building, indexing, or storing to lists.

var foo = [1, 1 + 1];
print foo[add(0, 1)]; // Expect: 2

foo[foo[0]] = 7;
print foo[3 - 2]; // Expect: 7

Now let’s make the list a bit more dynamic. We should be able to append items to the end of the list. Also, there should be a way to delete items at a specific index. In the name of simplicity, I added this functionality as native functions. Some languages — Go as an example — do this too. However, most languages expose this functionality in other ways. More on that later.

var foo = [1, 2, 3];

// append returns nil because it modifies foo in place
print append(foo, 4); // Expect: nil
// Now foo = [1, 2, 3, 4]

// delete returns nil because it modifies foo in place
print delete(foo, 2); // Expect: nil
// Now foo = [1, 2, 4]

Great work, users can now grow and shrink lists as they wish. This concludes the list semantics that we will be implementing in this post. Again for simplicity’s sake, a number of common list features have been left out. These include slicing print foo[1:8];, reverse indexing print foo[-1];, and more. For more details on some of these items you can check out the challenges section

Building a Runtime

Now that we know what semantics lists should have, we need to update the interpreter’s runtime. This is where the rubber hits the road. The runtime is just some C code that does the work of executing our Lox source code. If the runtime for lists is slow, then you can be sure that lists will be slow in Lox too. Considering this, how should we implement the runtime for lists? Two data structures immediatley come to mind: dynamic arrays and linked lists.

Let’s examine linked lists first. This data structure consists of nodes holding both data and a pointer to the next node in the list. You may recall that a linked list was used in the implementation of Clox to link all the objects together for garbage collection. Leaving the formal proofs for another time a linked list has the following algorithmic complexities; O(n) to access or modify an item; O(1) to append an item to the list; O(n) to delete an item.

O(n) complexity to access or modify an item presents an issue. We’d like something as common as accessing an item to be faster than that. The final nail in the coffin for linked lists is that the nodes aren’t guaranteed to exist in contiguous sections of memory. Scattered memory means terrible cache locality and as a result slow speeds.

Maybe dynamic arrays will have more favorable characteristics. Spoiler alert, they do. Dynamic arrays have been used all across the implementation of Clox. Thankfully the implementation is simple. It’s nothing more than a standard array with some additional metadata tracking when the array should grow. To grow the array simply allocate a larger block of memory and copy the values over. For algorithmic complexities we get; O(1) to access or modify an item; O(1) (amortized time) to append an item to the end of the list; O(n) to delete an item. Finally, because an array is a section of contiguous memory we see much better cache locality. All things considered, a dynamic array is a better fit for our list implementation.

To make the dynamic array cooperate with the rest of Clox we will define a new ObjList struct that extends the base obj struct in object.h. Remember that in the Clox implementation an object is a subtype of a Clox value. Other examples of objects include strings, functions, etc.

typedef struct {
    Obj obj;
    int count;
    int capacity;
    Value* items;
} ObjList;

Now let’s add the associated functionality of lists to object.c. We need functions for retrieving, storing, appending, and deleting items in the list, a function to build a new list for us, and a helper function to verify indices. It’s a lot of code, but keep in mind we’re really just writing a wrapper around a dynamic array.

ObjList* newList() {
    ObjList* list = ALLOCATE_OBJ(ObjList, OBJ_LIST);
    list->items = NULL;
    list->count = 0;
    list->capacity = 0;
    return list;
}

void appendToList(ObjList* list, Value value) {
    // Grow the array if necessary
    if (list->capacity < list->count + 1) {
        int oldCapacity = list->capacity;
        list->capacity = GROW_CAPACITY(oldCapacity);
        list->items = GROW_ARRAY(list->items, Value, oldCapacity, list->capacity);
    }
    list->items[list->count] = value;
    list->count++;
    return;
}

void storeToList(ObjList* list, int index, Value value) {
    list->items[index] = value;
}

Value indexFromList(ObjList* list, int index) {
    return list->items[index];
}

void deleteFromList(ObjList* list, int index) {
    for (int i = index; i < list->count - 1; i++) {
        list->items[i] = list->items[i+1];
    }
    list->items[list->count - 1] = NIL_VAL;
    list->count--;
}

bool isValidListIndex(ObjList* list, int index) {
    if (index < 0 || index > list->count - 1) {
        return false;
    }
    return true;
}

Still with me? Fantastic, because we just implemented the entire runtime for lists. Not too shabby. Note that if you are following along with your own implementation there are still a few things you’ll need to add. Notably, type checking macros like IS_LIST/AS_LIST, function declarations in object.h, and code to print our new list object.

Operation Opcode

Clox is a bytecode interpreter. This means that it compiles source code into a stream of bytecode which the VM can then execute via the runtime. Bytecode consists of some opcodes (1-byte instructions) with optional operands (1-byte arguments). To implement lists we will need to add some new opcodes to Clox.

The key to designing opcodes for a new feature is to K.I.S.S. — Keep It Simple Stupid. That is to say, if the language’s current set of opcodes can already support the new semantics you want, then do not add more opcodes. If you aren’t that lucky, then try to find the minimal number of opcodes you can add to support the new semantics. Minimizing the number of different opcodes is a tradeoff. It simplifies the VM implementation and helps you keep the common case fast. However, this comes at the cost of more complicated parsing and more time spent decoding bytecode. All the new semantics we are adding can be achieved by three new opcodes: OP_BUILD_LIST, OP_INDEX_SUBSCR, and OP_STORE_SUBSCR.

Many parts of this design and implementation are heavily influenced by Python. This includes both the semantics and syntax. The opcodes are no exception and mirror the opcodes Python uses for its list type.

OP_BUILD_LIST does the obvious thing. Notably, the opcode takes an operand itemCount. This is the number of values on the stack that it should build into the list. To ensure the proper order of the list, we peek at the items on the stack in reverse order4. Then we pop them all off the stack. Finally, it pushes the new list onto the stack. Here is the switch-case for it in vm.c:

case OP_BUILD_LIST: {
    // Stack before: [item1, item2, ..., itemN] and after: [list]
    ObjList* list = newList();
    uint8_t itemCount = READ_BYTE();

    // Add items to list
    push(OBJ_VAL(list)); // So list isn't sweeped by GC in appendToList
    for (int i = itemCount; i > 0; i--) {
        appendToList(list, peek(i));
    }
    pop();

    // Pop items from stack
    while (itemCount-- > 0) {
        pop();
    }

    push(OBJ_VAL(list));
    break;
}

OP_INDEX_SUBSCR stands for index subscript — meaning access an item at a particular index. It’s corresponding code in vm.h looks complex, but this is only because of a plethora of error handling. All it actually does is pop a list and index off of the stack, run indexFromList and push the resulting value back on the stack.

case OP_INDEX_SUBSCR: {
    // Stack before: [list, index] and after: [index(list, index)]
    Value index = pop();
    Value list = pop();
    Value result;

    if (!IS_LIST(list)) {
        runtimeError("Invalid type to index into.");
        return INTERPRET_RUNTIME_ERROR;
    }
    ObjList* list = AS_LIST(list);

    if (!IS_NUMBER(index)) {
        runtimeError("List index is not a number.");
        return INTERPRET_RUNTIME_ERROR;
    }
    int index = AS_NUMBER(index);

    if (!isValidListIndex(list, index)) {
        runtimeError("List index out of range.");
        return INTERPRET_RUNTIME_ERROR;
    }

    result = indexFromList(list, AS_NUMBER(index));
    push(result);
    break;
}

OP_STORE_SUBSCR of course means store subscript. Instead of pulling a value from the list, we are storing a value to the list. Again, error handling makes a simple task look complicated. First pop a list, index, and item from the stack. Then store the item in the list at the particular index.

case OP_STORE_SUBSCR: {
    // Stack before: [list, index, item] and after: [item]
    Value item = pop();
    Value index = pop();
    Value list = pop();

    if (!IS_LIST(list)) {
        runtimeError("Cannot store value in a non-list.");
        return INTERPRET_RUNTIME_ERROR;
    }
    ObjList* list = AS_LIST(list);

    if (!IS_NUMBER(index)) {
        runtimeError("List index is not a number.");
        return INTERPRET_RUNTIME_ERROR;
    }
    int index = AS_NUMBER(index);

    if (!isValidListIndex(list, index)) {
        runtimeError("Invalid list index.");
        return INTERPRET_RUNTIME_ERROR;
    }

    storeToList(list, index, item);
    push(item);
    break;
}

And that concludes designing and implementing the new opcodes required for lists in Lox. For a full implementation be sure to update debug.c with switch-cases for the new opcodes.

The Power of Parsing

Up until this point, our interpreter still can’t handle lists end to end. Hypothetically, the interpreter could execute hand compiled bytecode. But, hand compiling is no fun, so let’s automate it.

First, we need to get more formal about our syntax’s grammar. I’ve shown an excerpt of the grammar below that includes the modifications we will be making.

...

assignment     → ( call "." )? IDENTIFIER ( "[" logic_or "]" )* "=" assignment
               | logic_or ;

...

call           → subscript ( "(" arguments? ")" | "." IDENTIFIER )* ;
subscript      → primary ( "[" logic_or "]" )* ;
primary        → "true" | "false" | "nil" | "this"
               | NUMBER | STRING | IDENTIFIER | "(" expression ")"
               | "super" "." IDENTIFIER | "[" list_display? "]" ;

list_display   → logic_or ( "," logic_or )* ( "," )? ;

...

The biggest change is that we have added a list literal to the primary rule. It uses the utility rule list_display which accepts multiple comma-separated items. The items are any valid expression except for an assignment as indicated by the rule logic_or. list_display also supports an optional trailing comma as we wanted with the snippet ( "," )?.

To support indexing from lists, we’ve added a new rule subscript. It has a higher precedence than call. This allows things like indexing into a list for a function and then calling it e.g. myFunctions[0](). For storing to lists we updated the assignment rule. Both when accessing and storing items in a list, a valid index is any expression except an assignment again indicated by the rule logic_or.

With a formal grammar in hand, we can tie this new feature into compiler.c. First, we add two new rules to the Pratt parsing table. When we encounter a [ in a prefix scenario we parse a list literal. Seeing an infix [ kicks off parsing of a subscript (index or store).

// {Prefix, Infix, Precedence}
{ list, subscript, PREC_SUBSCRIPT }, // TOKEN_LEFT_BRACKET
{ NULL, NULL,      PREC_NONE },      // TOKEN_RIGHT_BRACKET

Below is the code to parse and compile a list literal. The left bracket is already consumed by the time it enters the list rule. If it doesn’t immediately find a right bracket it starts parsing comma-separated items one at a time. There is also an if statement to ensure it doesn’t try to parse an item after a trailing comma. Finally, it emits an OP_BUILD_LIST and the itemCount operand.

static void list(bool canAssign) {
    int itemCount = 0;
    if (!check(TOKEN_RIGHT_BRACKET)) {
        do {
            if (check(TOKEN_RIGHT_BRACKET)) {
                // Trailing comma case
                break;
            }

            parsePrecedence(PREC_OR);

            if (itemCount == UINT8_COUNT) {
                error("Cannot have more than 256 items in a list literal.");
            }
            itemCount++;
        } while (match(TOKEN_COMMA));
    }

    consume(TOKEN_RIGHT_BRACKET, "Expect ']' after list literal.");

    emitByte(OP_BUILD_LIST);
    emitByte(itemCount);
    return;
}

The code is simpler for the subscript rule. First, it parses the index via parsePrecedence(PREC_OR). Next, it emits an OP_STORE_SUBSCR if it sees an equal sign. Otherwise, it emits OP_INDEX_SUBSCR.

static void subscript(bool canAssign) {
    parsePrecedence(PREC_OR);
    consume(TOKEN_RIGHT_BRACKET, "Expect ']' after index.");

    if (canAssign && match(TOKEN_EQUAL)) {
        expression();
        emitByte(OP_STORE_SUBSCR);
    } else {
        emitByte(OP_INDEX_SUBSCR);
    }
    return;
}

Naturally, the parser expects that the source code has already been tokenized by the scanner. To make lists work we need to handle the cases for two new tokens. Thanks to the good design of Clox adding this is trivial.

case '[': return makeToken(TOKEN_LEFT_BRACKET);
case ']': return makeToken(TOKEN_RIGHT_BRACKET);

Finishing Up

Alright, let’s come up for air. That was a lot of code. We should be proud of ourselves though, we finally have lists working end to end. Unfortunately, we are still missing a few key parts of the implementation before we can call it done. We need to do two things: implement the append and delete functions, and update the garbage collector.

Append and Delete

append and delete will be available to the user through the native function interface. Both functions are simple wrappers around the runtime we built earlier in the post. The Clox native function interface does not currently handle errors so this has been left as an exercise to the reader. Here is the code in vm.c.

static Value appendNative(int argCount, Value* args) {
    // Append a value to the end of a list increasing the list's length by 1
    if (argCount != 2 || !IS_LIST(args[0])) {
        // Handle error
    }
    ObjList* list = AS_LIST(args[0]);
    Value item = args[1];
    appendToList(list, item);
    return NIL_VAL;
}

static Value deleteNative(int argCount, Value* args) {
    // Delete an item from a list at the given index.
    if (argCount != 2 || !IS_LIST(args[0]) || !IS_NUMBER(args[1])) {
        // Handle error
    }

    ObjList* list = AS_LIST(args[0]);
    int index = AS_NUMBER(args[1]);

    if (!isValidListIndex(list, index)) {
        // Handle error
    }

    deleteFromList(list, index);
    return NIL_VAL;
}

Garbage Collection

Ensuring that garbage collection was working correctly was the most difficult part of the implementation. On its face, it is a relatively simple task. To wire up lists to the garbage collector we need to handle two switch-cases in memory.c. In blackenObject we mark every item in the list.

case OBJ_LIST: {
    ObjList* list = (ObjList*)object;
    for (int i = 0; i < list->count; i++) {
        markValue(list->items[i]);
    }
    break;
}

In freeObject we want to free both the dynamic array and the object itself.

case OBJ_LIST: {
    ObjList* list = (ObjList*)object;
    FREE_ARRAY(Value*, list->items, list->count);
    FREE(ObjList, object);
    break;
}

With garbage collection, the devil is in the details. After lots of testing and careful inspection of the code, I found two bugs in my implementation. The fixes have already been included in the code of this post but I’ll outline the bugs for your benefit.

  1. In the OBJ_LIST switch-case of freeObject, I was only freeing the dynamic array. By forgetting to free the object itself too I had introduced a memory leak.

  2. You may have noticed the peculiar pop and push statements in the code for OP_BUILD_LIST. Those are there because appendList allocates memory which may kick off a garbage collection cycle. Since list isn’t rooted anywhere it would get swept away. Pushing it onto the stack prevents that from happening.

push(OBJ_VAL(list)); // So list isn't sweeped by GC in appendToList
for (int i = itemCount; i > 0; i--) {
    appendToList(list, peek(i));
}
pop();

All Done

Congratulations! We’ve finished a complete implementation of a performant list type. This is no small feat. Lists will certainly make programming in Lox much nicer. I hope that working through this encourages you to go and hack on programming languages too. Thanks for reading!

If you are looking for more things to explore beyond what we’ve gone over in this post then you’re in luck. Following in the footsteps of the book Crafting Interpreters I’ve included some challenges below.

Challenges

  1. More fully-featured languages often present a wider variety of ways to access items in a list. This includes negative indexing and slicing. Negative indexing is quite simple; an index of -1 accesses the last item, -2 the second last, and so forth. Slicing allows a user to easily extract and operate on portions of a list. In Python, something like myList[2:8:2] would take every second item of the list starting at index 2 and going to index 8 exclusive. Try supporting negative indexing in OP_INDEX_SUBSCR. Then try adding a native function with the signature Value slice(start, stop, step).

  2. Making append and delete native functions kept the implementation simple, but is uncommon in other languages. Two other options exist. Append and delete could be keywords that form a statement e.g. delete foo[0]. This statement would then be compiled down into some new opcodes. Alternatively, append and delete could be made into methods on a list object. This would require some rewiring but would perhaps be more idiomatic. Pick the approach you prefer and try implementing it.

  3. Most languages have a unified theory on iterable types. This includes how you work with lists/arrays, iterators, strings, generators, and more. Currently, our implementation is pretty lacking in this area. Research how other languages implement iterators and try adding them to Clox. Additionally, try adding the ability to index individual characters of a Clox string.

  4. The current implementation of delete only removes an item and reshuffles the content in the array as necessary. If many items were deleted from a list then the list would be overallocated for memory. Try adding some logic to delete that deallocates memory when the number of items in the list becomes small enough.

Leave a comment on Reddit / Hacker News


  1. Lots of code will be shown but it will not be comprehensive. Not every line of code required for a complete implementation will be provided. ↩︎

  2. Lox is a dynamic programming language. In Crafting Interpreters you build two interpreters for the language. Jlox a tree-walk interpreter with Java and Clox a bytecode interpreter with C. This post modifies the Clox interpreter. I originally developed the code for adding lists to Lox on my own Lox derivative called NQQ. ↩︎

  3. In the Clox implementation parsing and compilation are squeezed into a single step. Despite the fact that they are happening at the same time, they are still providing different functions. Parsing is all about turning a flat stream of tokens into a hierarchical structure that represents the intent of the computation. Compilation is about turning that structure into something that is easier and quicker to execute. By squishing these two steps into one we are skipping building the AST. This has the possible benefits of being conceptually simpler and faster but inhibits static analysis of the code. ↩︎

  4. When a list is parsed the items are pushed onto the stack one at a time from left to right. If OP_BUILD_LIST were to read the items in from the top of the stack to the bottom it would be building the list in reverse. ↩︎