D for the Win
August 21, 2014

I'm a convert! I've seen the light!

By the way, be sure to read part 2 as well.

You see, Python is nice and all and it excels in so many domains, but it was not crafted for the ever growing demands of the industry. Sure, you can build large-scale projects in Python (and I have built), but you take it out of the lab and into the real world, the price you pay is just too high. Literally. In terms of work per CPU cycle, you can't do worst.

The C10M problem is a reiteration of the C10K problem. In short, today's commodity hardware can handle millions of packets per second, but in reality you hardly ever reach such numbers. For example, I worked a short while at a company that used AWS and had tens of twisted-based Python servers accepting and logging requests (not doing any actual work). They managed to squeeze ~500 requests/sec out of this setup (per machine), which escalated in cost rather quickly. Moving to PyPy (not without trouble) did triple the numbers or so, but still, the cost simply didn't scale.

Python, I love you, but you help instill Gate's law -- "The speed of software halves every 18 months". In the end, we pay for our CPU cycles and we want to maximize our profit. It's not you, Guido, it's me. I've moved on to the C10M world, and for that I'd need a programming language that's designed for system programming with a strong and modern type system (after all, I love duck typing). I need to interface with external systems, so a C ABI is desirable (no foreign function interface), and meta-programming is a huge plus (so I won't need to incorporate cumbersome code-generation in my build system). Not to mention that mission-critical code can't allow for the occasional NameError or NoneType has no member __len__ exceptions. The code must compile.

I've looked into rust (nice, but will require a couple of years to mature enough for a large-scale project) and go (Google must be joking if they actually consider it for system programming), but as strange as it may sound, I've finally found what I've been looking for with D.

Dlang Dlang Über Alles

System programming is a vast ocean of specifics, technicalities and constraints, imposed by your specific needs. Instead of boring you to death with that, I thought it would be much more intriguing to compare D and Python. In other words, I'll try to show how D speaks fluent Python.

But first things first. In (the probable) case you don't know much D -- imagine it's what C++ would have dreamed to be. It offers cleaner syntax, much shorter compilation time, (optional) garbage collection, highly expressive templates and type inference, Pythonic operator overloading (implemented as rewriting), object-oriented and functional capabilities (multi-paradigm like Python), intermingles high-level constructs (like closures) with low-level ones (naked functions in inline assembly) to produces efficient code, has strong compile-time introspection capabilities and some extra cools features in the domain of code generation: mixin -- which evaluates an arbitrary string of D code at compile time, and CTFE -- compile-time function execution. Whoa, that was long.

In general, D follows Python's duck-typed (or protocol-oriented) spirit. If a type provides the necessary interface ("protocol") for an operation, it will just work, but you can also test for compliance at compile time. For example, ranges are a generalization of generators in Python. All you need to do in order to be an InputRange is implement bool empty(), void popFront() and auto front(), and you can use isInputRange!T to test whether T adheres the protocol. By the way, the exclamation point (!), which we'll soon get acquainted with, distinguishes compile-time arguments from runtime ones.

For brevity's sake, I'm not going to demonstrate all the properties I listed up there. Instead, I'll show why Python programmers ought to love D.

Case Study #1: Generating HTML

In an old blog post I outlined my vision of HTML templating languages: kill them all. I argued they are all but crippled-down forms of Python with an ugly syntax, so just give me Python and an easy way to programmatically manipulate the DOM.

I've later extended the sketch into a library in its own right, named srcgen. You can use it to generate HTML, C-like languages and Python/Cython code. I used it in many of my commercial projects when I needed to generate code.

So here's an excerpt of how's it done in srcgen:

def buildPage():
    doc = HtmlDocument()
    with doc.head():
        doc.title("das title")
        doc.link(rel = "foobar", type="text/css")

    with doc.body():
        with doc.div(class_="mainDiv"):
            with doc.ul():
                for i in range(5):
                    with doc.li(id = str(i), class_="listItem"):
                        doc.text("I am bulletpoint #", i)

    return doc.render()

And here's how it's done in D:

auto buildPage() {
    auto doc = new Html();

    with (doc) {
        with (head) {
            title("das title");
            link[$.rel = "foobar", $.type = "text/css"];
        }
        with (body_) {
            with(div[$.class_ = "mainDiv"]) {
                with (ul) {
                    foreach(i; 0 .. 5) {
                        with (li[$.id = i, $.class_ = "listItem"]) {
                            text("I am bulletpoint #");
                            text(i);
                        }
                    }
                }
            }
        }
    }

    return doc.render();
}

You can find the source code on github, just keep in mind it's a sketch I wrote for this blog post, not a feature-complete library.

The funny thing is, Python's with and D's with are not even remotely related! The Python implementation builds a stack of context managers, while with in D merely alters symbol lookup. But lo and behold! The two versions are practically identical, modulo curly braces. You get the same expressive power in both.

Case Study #2: Construct

But the pinnacle is clearly my D version of Construct. You see, I've been struggling for many years to create a compiled version of Construct. Generating efficient, static code from declarative constructs would make the library capable of handling real-world data, like packet sniffing or processing of large files. In other words, you won't have to write a toy parser in Construct and then rewrite it (by hand) in C++.

The issues with my C version of Construct were numerous, but they basically boiled down to the fact I needed a stronger object model to represent strings, dynamic arrays, etc., and adapters. The real power of Construct comes from adapters, which operate at the representational ("DOM") level of the data, rather on its binary form. That required lambdas, closures and other higher-level concepts that C lacks. I even tried writing a Haskell version, given that Haskell is so high-level and functional, but my colleague and I had given hope after a while.

Last week, it struck me that D could be the perfect candidate: it has all the necessary high-level concepts while being able to generate efficient code with meta-programming. I began fiddling with a D version, which proved extremely promising. So without further ado, I present dconstruct -- an initial sketch of the library.

This is the canonical PascalString declaration in Python:

>>> pascal_string = Struct("pascal_string",
...     UBInt8("length"),
...     Array(lambda ctx: ctx.length, Field("data", 1),),
... )
>>>
>>> pascal_string.parse("\x05helloXXX")
Container({'length': 5, 'data': ['h', 'e', 'l', 'l', 'o']})
>>>
>>> pascal_string.build(Container(length=5, data="hello"))
'\x05hello'

And here's how it's done in D:

struct PascalString {
    Field!ubyte length;
    Array!(Field!ubyte, "length") data;

    // the equivalent of 'Struct' in Python,
    // to avoid confusion of keyword 'struct' and 'Struct'
    mixin Record;
}

PascalString ps;
auto stream = cast(ubyte[])"\x05helloXXXX".dup;
ps.unpack(stream);
writeln(ps);
// {length: 5, data: [104, 101, 108, 108, 111]}

Through the use of meta-programming (and assuming inlining and optimizations), that code snippet there actually boils down to something like

struct PascalString {
    ubyte length;
    ubyte[] data;

    void unpack(ref ubyte[] stream) {
        length = stream[0];
        stream = stream[1 .. $]; // advance stream
        data = stream[0 .. length];
        stream = stream[length .. $];  // advance stream
    }
}

Which is as efficient as it gets.

But wait, there's more! The real beauty here is how we handle the context. In Python, Construct builds a dictionary that travels along the parsing/building process, allowing constructs to refer to previously seen objects. This is possible in D too, of course, but it's highly inefficient (and not type safe). Instead, dconstruct uses a trick that's commonly found in template-enabled languages -- creating types on demand:

struct Context(T, U) {
    T* _curr;
    U* _;
    alias _curr this;   // see below
}

auto linkContext(T, U)(ref T curr,  ref U parent) {
    return Context!(T, U)(&curr, &parent);
}

The strange alias _curr this is a lovely feature of D known as subtyping. It basically means that any property that doesn't exist at the struct's scope will we forwarded to _curr, e.g., when I write myCtx.foo and myCtx has no member named foo, the code is rewritten as myCtx._curr.foo.

As we travel along constructs, we link the current context with its ancestor (_). This means that for each combination of constructs, and at each nesting level, we get a uniquely-typed context. At runtime, this context is nothing more than a pair of pointers, but at compile time it keeps us type-safe. In other words, you can't reference a nonexistent field and expect the program to compile.

A more interesting example would thus be

struct MyStruct {
    Field!ubyte length;
    YourStruct child;

    mixin Record;
}

struct YourStruct {
    Field!ubyte whatever;
    Array!(Field!ubyte, "_.length") data;  // one level up, then 'length'

    mixin Record;
}

MyStruct ms;
ms.unpack(stream);

When we unpack MyStruct (which recursively unpacks YourStruct), a new context ctx will be created with ctx._curr=&ms.child and ctx._=&ms. When YourStruct refers to "_.length", the string is implanted into ctx, yielding ctx._.length. If we refered to the wrong path or misspelled anything, it would simply not compile. That, and you don't need dictionary lookups at runtime -- it's all resolved during compilation.

So again, this is a very preliminary version of Construct, miles away from production grade, but you can already see where it's going.

By the way, you can try out D online at dpaste and even play around with my demo version of dconstruct over there.

In Short

Python will always have a special corner in my heart, but as surprising as it may be (for a guy who's made his career over Python), this rather unknown, rapidly-evolving language, D, has become my new language of choice. It's expressive, concise and powerful, offers short compilation times (as opposed to C++) and makes programming both fun and efficient. It's the language for the C10M age.