In discussions on programming languages I often see C being designated as a neat, successful language that makes the right tradeoffs. People will go so far as to say that it's a "small language", it "fits in your head" and so on.

I can only imagine that people saying these things have forgotten how much effort it was to really learn C.

I've seen newbies ask things like "I'm a java coder, what book should I use to learn C?" And a lot people will answer K&R. Which is a strange answer, because K&R is a small book (to further perpetuate this idea that it's a small language), is not exactly pedagogical, and still left me totally confused about C syntax.

In practice, learning C takes so much more than that. If you know C the language then you really don't know anything yet.

Because soon enough you discover that you also need to know the preprocessor and macros, gcc, the linker, the loader, make and autoconf, libc (at least what is available and what is where - because it's not organized terribly well), shared libraries and stuff like that. Fair enough, you don't need it for Hello World, but if you're going to do systems programming then it will come up.

For troubleshooting you also need gdb and basically fundamental knowledge of your machine architecture and its assembly language. You need to know about memory segments and the memory layout and alignment of your datastructures and how compiler optimizations affect that. You will often use strace to discover how the program actually behaves (and so you have to know system calls too).

Much later, once you've mastered all that, you might chance upon a slide deck like Deep C whose message basically is that you don't understand anything yet. What's more terrifying is that the fundamental implication at play is: don't trust the abstractions in the language, because when things break you will need to know how it works under the hood.

In a high level language, given effort, it's possible to design an API that is easy to use and hard to misuse and where doing it wrong stands out. Not so in C where any code is always one innocuous looking edit away from a segfault or a catastrophic security hole.

So to know C you need all of that. But that's mostly the happy path. Now it's time to learn about everything that results in undefined behavior. Which is the 90% of the iceberg below the surface. Whenever I read articles about undefined behavior I'm waiting for someone to pinch me and say the language doesn't actually allow that code. Why would "a = a++;" not be a syntax error? Why would "a[i]" and "i[a]" be treated as the same when syntactically they so clearly aren't?

Small language? Fits in your head? I don't think so.

Oh, and once you know C and you want to be a systems programmer you also need to know Posix. Posix threads, signals, pipes, shared memory, sync/async io, ... well you get the idea.

I'm inspired by how many great Python libraries there are these days, and how easy it is to use them. requests is the canonical example, and marks a real watershed moment, but there are many others.

It made me think back on various projects that I've published over the years and not touched in ages. I've been considering them more or less "complete". My standards for publishing projects used to be: write a blog entry, include the code, done. That was okay for simple scripts. Later on I started putting code on berlios.de and sourceforge.net. At some point github emerged and became the de facto standard, so I started using that too.

Fast forward to 2014 and the infrastructure available to open source projects has been greatly enriched. And with it, the standards for what makes a decent project have evolved. Jeff Knupp wrote a fabulous guide on this.

I decided to pick a simple case study. ansicolor is a single module whose origins I can trace back to 2008. I've seen the core functionality present in any number of codebases, because it's just so easy to hammer out some code for this and call it a day. But I never found it in a reusable form, so I decided to make it a separate thing that I could at least reuse between my own projects.

These are the steps a project is destined to pass through:

• python3 support
• pypi package + wheel!
• readme that covers installation and "getting started"
• tests + tox config
• travis-ci hook
• flake8 integration and fixing style violations
• docs + Read the Docs hook

Not a single feature was added to ansicolor, not a single API was changed. Only two things really changed at the level of the code: exports were tidied up and docstrings were added. Python3 support was added too, but it was so trivial you'd have to squint to notice it.

The biggest stumbling block was actually writing the docs. As an implementor you tend to look at code in a completely different light than you do as a user of that code. Before starting on this I was thinking about how the API is a bit awkward in some places and could be improved. And how some of the functionality caters to a very narrow use case and maybe should be removed or to moved to a "contrib"-like place.

But as a potential user of a library that I just discovered I don't care about any of that. I want to be able to "pip install" it. I want to have some quickstart documentation so I can have running code in 2 minutes. That's how long I'll typically spend deciding whether this code is worth my time at all, so if the implementor is busy polishing the API before even putting out a pypi package they're wasting their time.

There is an interesting cognitive dissonance at play here. As an implementor I tend to think that the darkest corners of my code are those that most need documenting. Those are the ones most likely to bite someone. The easy stuff anyone can figure out. But as a user that's not how I see it at all. It's precisely the simplest functionality that most needs explaining, because most users have simple needs. If you do a good job documenting that you can make lots of people productive. By contrast, the complicated features have a small audience. An audience that's more sophisticated and more likely to help themselves by reading the code if need be.

Then there are the tools. I always found sphinx a bit fiddly. It's not really obvious how to get what you want, and it's permissive enough not to complain, so it takes a fair bit of doc hunting to discover how other projects do it. PyPI has a more conservative rst parser than github, so if you give it syntax it doesn't accept it renders your page in plain text. I ended up doing a number of releases where only the readme changed slightly to debug this. Read the Docs works well, but I couldn't figure out how to make it build from a development branch. It seems to only want to build from a tag regardless of the branches you select, so that too inflated the number of releases.

It takes a bit of time to renovate a project, but it's all fairly painless. All these tools have reached a level of maturity that makes them very nice to use.

So, writing an interpreter. Where do we even start?

Why, with the essentials. While an interpreter is nothing more than a program that reads your program and runs it, that is a pretty complicated undertaking. You could conceivably write an interpreter that reads a program character by character and tries to execute it, but you would soon run into all kinds of problems, like: I need to know the value of this variable, but that's defined somewhere else in the program, and: I'm executing a function definition, which doesn't have any effect, but I need to remember that I've seen it in case anyone is calling the function later in the program.

In fact, this approach is not wrong, it's basically what your machine really does. But your machine is running machine code, and you're not, and therein lies the difference. People don't write programs in machine code anymore. It's too hard. We have high level languages that provide niceties like variables - define once, reference as often as you like. Like functions. Like modules. All of these abstractions amount to a program that is not a linear sequence of instructions, but a network of blocks of instruction sequences and various language specific structures like functions and classes that reference each other.

And yet, since many languages are compiled down to machine code, there must be a way to get from here to there? Well, compiler writers have decided that doing so in one step is needlessly complicated. And instead we do what we always do in software engineering when the problem is too hard: we split it into pieces. A compiler will run in multiple phases and break down the molecule into ever simpler compounds until it's completely flat and in the end you can execute it sequentially.

Interpreters start out the same way, but the goal is not machine code, it's execution of the program. So they bottom out in an analogous language to machine code: byte code. Byte code can be executed sequentially given a virtual machine (language specific) that runs on top of a physical machine (language agnostic) that knows a few extra things about the language it's running.

Phases

Parsing - Read the program text and build a structured representation (abstract syntax tree) out of it.

Compilation - Compile the AST to bytecode instructions that can be executed on a stack machine.

Interpretation - Pretend to be a machine and execute the bytecode.

The parser is the thing that will read the program text, but it needs to know the syntax of the programming language to recognize what it's reading. This is supplied in the form of a grammar. The parser also needs to know the structure of your abstract syntax for it to build those trees for you. And this is where it gets tricky, because the AST is one of the central data structures in an interpreter, and the way you design the AST will have a big impact on how awkward or convenient it will be to do anything meaningful with it. Even so, the AST much match the grammar very closely, otherwise it's going to be very hard to write that parser. So the optimal would be to design the AST first, then write a grammar that matches it to the letter (even better: you could generate the grammar from the AST). Well, there are decades of research that explain why this is hard. Long story short: the more powerful the parser the more leeway you have in the grammar and the easier it is to get the AST you want. Chances are you will not write the parser from scratch, you'll use a library.

The compiler is the thing that traverses an AST and spits out bytecode. It's constrained by two things: the structure of the AST, and the architecture of the vm (including the opcodes that make up your bytecode). Compilers really aren't generic - they need to know all the specifics of the machine they are compiling for. For virtual machines this means: is it a stack machine or a register machine? What are the opcodes? How does the vm supply constants and names at interpretation time? How does it implement scope and modules?

Finally, the interpreter is the thing that emulates a machine and executes the bytecode. It will read the bytecode sequentially and know what to do with each instruction. It's really the only part of the entire program that interacts with its environment extensively. It's the thing that produces output, opens sockets, creates file descriptors, executes processes etc, all on behalf of your program. In fact, it would be more correct to say that these things are contained in the standard libraries of the language, but the interpreter is what orders them to happen.

luna 0.0.1

It's out. An interpreter for basic blocks. All your code runs in a single stack frame, so you can only run expressions, assignments and function calls. There is an object model for lua values. There is a pretty barebones standard library. And there are tests (py.test is making life quite good so far).

• luna

I've been doing a bunch of reading around languages and interpreters lately, but I want to convert that insight into the kind of knowledge that only comes from the experience of implementation. I feel woefully unprepared for this, but on the other hand you can never prepare sufficiently for anything in life with research alone. I expect to hit many stumbling blocks, but that's kind of the point - without hitting them I wouldn't learn much.

Why lua?

Lua as a language has some interesting and unusual properties:

• Designed to be small.
• Designed to be portable.
• Designed to be embeddable.

This is not just aspirational, lua really delivers on these promises.

How small?

• the grammar has 21 keywords and a small set of productions,
• 8 types: nil, boolean, number (double precision float), string, function, userdata, thread, and table,
• a single kind of data structure: tables (associative arrays),
• no type declarations (dynamically typed),
• no manual memory management (garbage collected),
• no object system,
• no exceptions.

How portable? The lua interpreter is written in ANSI C. The implementors have consistently snubbed platform-specific services so that it builds just about anywhere, and from a single Makefile at that. That's right, no autotools. Try that with gcc.

How embeddable? In a word: very. The interpreter is only 15k sloc. The binary takes all of 6 seconds to build on my platform (linux x86, lua 5.2.2) and is 211kb. Lua is used more in embedded form that standalone and it's embedded in an enormous amount of software. Being small and portable makes it a great language to embed, because it's easy to learn and easy to use.

What's luna?

It's a little known fact that Lua, who is Brazilian, has an Italian sister called Luna. It's not hard to see why that would be - Luna has very little in common with her sister. She doesn't like to travel (not portable) and she doesn't like crowds (not embeddable). Luna has none of the qualities that make Lua so popular.

Not only that, Luna is very careless. When she's doing a task she doesn't care how long she takes, because life is good, so why hurry?

Of course, Python is no language to write an interpreter in. You should be using RPython or C. But it's the perfect language to prototype designs in when you're not really sure what you're doing yet.

Roadmap (tentative)

• 0.0.1 - An interpreter that can execute a basic block. At this stage a program can consist of: assignment, expressions, function calls.
  • A parser and AST that cover all expressions and a small subset of statements.
  • A bytecode compiler.
  • A single stack frame virtual machine (stack based) with a single global environment.
• 0.0.2 - Blocks and control flow: if, for, while, repeat.
  • A vm with multiple environments that understands lexical scope.
• 0.0.3 - Function definitions.
  • A vm with multiple stack frames.

References

1. Lua 5.1 Reference Manual
2. The Implementation of Lua 5.0
3. Where Lua Is Used

New to the series? The previous entry was part 4.

We have seen quite a bit of bytecode already - we've even seen the structure of a bytecode module. But we haven't seen loops yet.

A while loop

We'll start with the simplest kind of loop. It's not really a typical loop given that it breaks after a single execution, but it's enough to give us a look at the looping machinery.

def loop(x):
    while x > 3:
        print x
        break

Disassembly of loop:
  2           0 SETUP_LOOP              22 (to 25)
        >>    3 LOAD_FAST                0 (x)
              6 LOAD_CONST               1 (3)
              9 COMPARE_OP               4 (>)
             12 POP_JUMP_IF_FALSE       24

  3          15 LOAD_FAST                0 (x)
             18 PRINT_ITEM          
             19 PRINT_NEWLINE       

  4          20 BREAK_LOOP          
             21 JUMP_ABSOLUTE            3
        >>   24 POP_BLOCK           
        >>   25 LOAD_CONST               0 (None)
             28 RETURN_VALUE

The control flow is a bit convoluted here. We start off with a SETUP_LOOP which pushes a block onto the stack. It's not really clear to me why that's necessary, given that Python does not use blocks for scopes. But it might be that the interpreter needs to know which level of looping it's at.

We then load the variable x and the constant 3 and run COMPARE_OP. This opcode actually takes a parameter to tell it which operation to perform (greater than in this case). The result of that will be a boolean value on the stack.

Now we need to know whether we're going to execute the loop body or jump past the loop, so that's POP_JUMP_IF_FALSE, which may jump to location 24 where the loop ends.

Assuming we are in the loop body, we simply load the variable x and print it. Interestingly, the print statement requires two opcodes PRINT_ITEM and then PRINT_NEWLINE, which seems a bit over the top.

We now have a BREAK_LOOP instruction. Notice that if we were to ignore and execute the JUMP_ABSOLUTE just behind it that would return us to the loop predicate, and we might continue looping. But that's not supposed to happen after a break: a break ends the loop even if the loop predicate is still true. So this must mean that we won't reach JUMP_ABSOLUTE.

After this we execute POP_BLOCK which will officially end the loop by taking the block off the stack again.

A for loop

A for loop, then, is not very different. The main difference is that we are not looping on a boolean condition - we are looping over an iterable.

def loop(x):
    for i in range(x):
        print(i)

Disassembly of loop:
  2           0 SETUP_LOOP              25 (to 28)
              3 LOAD_GLOBAL              0 (range)
              6 LOAD_FAST                0 (x)
              9 CALL_FUNCTION            1
             12 GET_ITER            
        >>   13 FOR_ITER                11 (to 27)
             16 STORE_FAST               1 (i)

  3          19 LOAD_FAST                1 (i)
             22 PRINT_ITEM          
             23 PRINT_NEWLINE       
             24 JUMP_ABSOLUTE           13
        >>   27 POP_BLOCK           
        >>   28 LOAD_CONST               0 (None)
             31 RETURN_VALUE

To do that we have LOAD_GLOBAL to load the range function on the stack. This is an opcode we haven't seeb before, and it simply means this name comes from somewhere outside this module (the __builtin__ module in this case). We then load x and call the function. This produces a list.

Now, since Python uses iterators so heavily, the loop will use this method to move through the list. It means you could also loop over any other iterable object (tuple, dict, string, your own custom iterators etc). In fact, GET_ITER amounts to calling the iter function on the list (which returns an iterator object). And FOR_ITER calls the iterator's next method to get the next item.

We now have the first int in the list, and we bind it to the name i with STORE_FAST. From there on, we may use i in the loop body.

You will notice that there is something odd about the way i is manipulated. At location 16 the int is sitting on the stack, and gets bound to a name with STORE_FAST. This consumes it on the stack. We then immediately push it on the stack again with LOAD_FAST. These two instructions cancel each other out: we could remove them without changing the meaning of the program.

So why do we have to store and load? Well, imagine i were used again in the loop body - it would have be bound, right? So it could be optimized away in this case, but not in the general case.