[Listening to: Aglaia – Three Organic Experiences. Track #3, Seven Ancient Glaciers is a definite stand-out!]
Welcome, stranger! This series is a journal that attempts to document my trials and tribulations as I embark on the impossible journey of finding the perfect programming language that I can use with great joy on my hobby projects. An epic quest of unfathomable difficulty, indeed!
In this episode: after a somewhat lacklustre introductory meandering on the merits of the multi-language approach in software development, things get heated up quite a bit with some good old-fashioned C++ bashing, after which the hero calms down and finally sets some GOALs. A rare Greek mythological creature also makes an appearance midway.
Performance vs productivity
A good programmer is a good programmer in any language. Yet the choice of language has significant effects on both developer productivity and the runtime performance of the resulting software. The “universally accepted truth” appears to be that these two things form a sort of a dichotomy. Low-level languages allow ultimate control over the hardware, thus yielding the best possible performance (assuming the programmer knows what he’s doing) and traversing up towards the high-level end of the scale languages gain progressively more expressive power which results in improved programmer productivity (more “impact” per lines of code), albeit at the cost of runtime speed. Programmer and program efficiency are widely believed to be mutually exclusive goals—you just can’t achieve both within the confines of a single language.
One solution to this problem has traditionally been the two-language approach, which boils down to combining a highly expressive but slow high-level language with a speedy but relatively primitive low-level one as an attempt to gain the best benefits of both worlds. There are generally two ways to go about this: by extending or embedding.
When extending, the bulk of the program is written in a high-level language (either interpreted or statically compiled) to get the benefit of higher programmer productivity, then performance critical parts simply get rewritten in a more efficient low-level language. Thus the high-level language gets extended with special purpose modules or libraries that handle the performance critical duties.
Certain languages, such as Python, are more suitable for extending than embedding. Arguably, one of the best mainstream examples for extending Python is the hugely popular sci-fi space MMORPG game EVE Online. The EVE client was written in Stackless Python that is extended with custom coded C++ modules to handle platform-specific graphics, sound, network I/O and other performance-sensitive tasks.
This method might work fine from a performance optimization point of view, but it’s not without drawbacks. First of all, there is very likely going to be an “impedance mismatch” between the two languages, which will necessitate to devise some sort of bridging mechanism to allow the interchange of data structures and bi-directional control flow between the two layers. This can quickly result in lots of glue code and data duplication, especially if the two languages were not designed to co-operate with each other. Also, the high-level language might have some quite rigid ideas about how the data should be laid out in memory that might further complicate the low-level performance optimization efforts (e.g. misaligned data, difficulty of achieving SoA layouts (Struct of Arrays) for SIMD optimizations etc.) While it’s certainly possible to speed the program up this way, the time gained by using a high-level language can easily be amortised by having to build a custom (and potentially suboptimal) bridging solution and the increased general complexity of the development workflow.
The second variant is writing a high-performance core in a low-level language (usually a statically compiled one, such as C or C++) and then embedding a high-level language into the program (typically an interpreted dynamic scripting language). The core functionality would be exposed through an API and the application logic would be implemented in the scripting language using this API. In this scenario, the high-level scripting language is relegated to an orchestrator role. Another related but slightly different usage pattern for embedding is to enable the end-user to extend the application via scripts and plugins, or in the case of games, to allow modding.
Lua is very well suited for embedding; over the years, it has almost become the de-facto scripting language of choice for C/C++ applications, especially since the advent of LuaJIT. Due to its very small memory footprint and high runtime efficiency, it has gained a widespread adoption in the C++ dominated gaming industry over the last two decades, and it has found its way into a wide variety of applications and embedded systems as well, such as high-performance web servers, in-memory databases, routers and audio software.
On the desktop application front, the best high-profile commercial example is undeniably Adobe Photoshop Lightroom. More than 60% of the Lightroom code was written in Lua (basically the whole UI logic), C++ was only used for the speed-critical image-manipulation routines and some platform-specific glue code.
Of course, the downside of the embedding approach is that a significant part of the program still has to be written in C/C++, which leads right into my next topic…
Life is too short for C++
I can almost hear some people murmuring in the back row at this point: “But we already have a language that combines ultimate bare-metal performance with high-level productivity: it’s called C++!”
Um, no. That couldn’t be further from the truth. For a start, read the preceding section again—if C++ was so awesome at everything, why would people want to embed high-level scripting languages into their applications? (Apart from plugin support and giving gameplay scripters a much simpler language to use, which are some very good reasons, but let’s focus on developer productivity here.)
C++ is a total disaster of a language, a bloated and overcomplicated monstrosity. Its only redeeming quality is its (almost) full backwards-compatibilty with C. That may sound overly harsh, but look at a few successful C++ open-source projects and check out some blog posts and rants written by AAA game developers who are forced to use this abomination on a daily basis. You’ll quickly see a common pattern emerging: the general best practice is to just stick to a very narrow, restricted subset of the language and stay away from its so-called “high-level abstractions”—the existence of which should normally be the main reason for choosing it over C in the first place.
So what happens with C++ on a successful project? The wretched standard library flies right out of the window (along with other ghoulish aberrations originating from the deepest bowels of hell like Boost), operator overloading is basically streng verboten at the risk of public corporal punishment (except for the simplest cases), template metaprogramming is best left untouched (but only if you value your sanity and want to get a working build out of the miserable compiler on the same day), so in the end what you’re left with is pretty much plain old C with some extra bells-and-whistles tacked on its back.
C++: an octopus made by nailing extra legs onto a dog.
And don’t even get me started with C++11—it’s a lesser known that what the “11” actually signifies in that name is the number of extra tentacles bolted to the language in an attempt to make it suck less hard…
So it’s a fair conclusion to say that using C++ in a sane way is using it as a C. Let’s remember that C is ultimately nothing more than a clever cross-platform assembler (plus a somewhat usable standard library). I do like C, it’s a nice minimal system-programming language that’s almost perfect at what it aims to do, and I have no problems using it for low-level work. But using it as a high-level language is just a futile exercise in frustration.
(Let me also quickly add that I have the utmost respect for all C++ coders out there who are capable of producing useful software with this beast of a language, despite all the odds (e.g. most game developers). Just to set the record straight, I am criticising the language, not its users.)
Um, is this the best we got?
Ok, back on track after this slight C++ bashing detour. The thing is, as long as C++ is still alive and well, I think it rightfully deserves all the bashing it can get. And if you really love it, congratulations, you’re one happy programmer, a content user of one of the most popular languages in existence on this planet today! That’s awesome, I wish you all the best, have fun and fare well!
Meanwhile, things are not looking exactly spectacular on our side. It seems that there’s nothing we can possibly do about this rather unfortunate situation if we want both ultimate programmer productivity and the best runtime efficiency, right? This is just the way it is, so we’d better suck it up and prepare to mount ourselves onto the back of our strange two-language mule. Being the astute reader of this fine piece of publication that you certainly are, you must have no doubt figured it out by now that there must be a better way…
Redefining the GOAL
One day I came across an interesting post on HackerNews about a most unusual language called GOAL that the game studio Naughty Dog created for the development of their older PlayStation titles. Now, GOAL stands for Game Oriented Assembly Lisp, and as the name implies, it allows the programmer to seamlessly intermix high-level Lisp code with down-to-the-metal assembly in the same lexical environment. Of course, it also supported all the killer features commonly associated with Lisp, such as inspecting and changing the program while it’s running, macros and all that sort of stuff (check out the code example in this forum post by one of the Naughty Dog devs).
I found this idea extremely cool! A language that integrates the two opposite extremes of the abstraction spectrum into a single coherent form! Assembly code would be written as S-expressions, so it could be generated using macros or manipulated as data just like Lisp code. No more messing around with language bridges, it’s all a single package.
The concept just got stuck in my head and I sort of got obsessed with the topic. I had been trying to seek out all information I could on GOAL, but sadly there was not too much to be found on the internet. The original sources are now owned by Sony, so it’s a safe bet that we’ll never get a chance to take a peek into them. Luckily, Andy Gavin, the genius who created GOAL had written up a quite nice overview on it which is definitely worth reading (for more interesting reading material, check out the reading suggestions at the end of this article).
It is also important to note that GOAL was not just simply a language but a whole interactive system that allowed rapid prototyping of ideas and exploratory programming, as described in this post by a veteran Naughty Dog developer.
Well, this all sounds very much what I’d like my ideal programming language to be: interactive and highly-expressive while not sacrificing runtime efficiency. So the question is, does such a language exist somewhere today, silently waiting to be discovered? We’ll find that out in the next part!