A-code language history (as I recall it)

The main purpose of this document is to explain (in response to some requests) why some aspects of my A-code implementation are the way they are, and what the future might hold if I get around to it.

A-code is the language developed by Dave Platt in order to write his influential Adv550 expansion of the classic game Adventure by Crowther and Woods (Adv350 in the modern nomenclature). Originally written in PL6, a Fortran 77 implementation of the original A-code engine was distributed in the late 1980s together with the the Adv550 A-code source.

Dave Platt's A-code engine had a "munger" and an "executive", both written in F77. The munger took the A-code source and produced a tokenised pseudo-binary. The executive was was effectively a virtual machine which executed the pseudo-binary. The lightly encoded game's text was in a separate file.

I first made use of this original A-code implementation to merge Luckett's and Pike's Adventure II (now known as Adv440) with Adv550 and Platt's Adv550 into Adventure4, which later evolved into Adventure4+ (now known as Adv660).

The initial Adventure4 and then Adventure4+ (1983 - 1985/6, on Primes) implementation took the A-code architecture as it was. The only changes to the munger/executive were in improving the command parser (e.g. automatically allowing all words to be abbreviated to the minimal unambiguous length, chaining commands on one line with semicolons, and providing AGAIN for command (simple or compound) repetition.

In September 1990 I embarked on re-implementing A-code in C on Unix, the main aim being to teach myself C (my previous expertise was first in Algol/Algol68, then in Fortran4, then in Fortran77). I later joked that the new version was dedicated to the proposition that a real programmer can write Fortran in any language. That implementation was the seed of the current version and it moved away from Platt's in one crucial step, the full significance of which did not transpire until much later. Instead of replicating Platt's virtual machine approach, I wrote acdc, which translated A-code directly into compilable C, with game-independent kernel C source providing a library of standard calls used by the translated code.

The reason for this shift was performance. Adventure4+ by that time became too large and complex, yet I wanted it to run on ordinary PCs as they then were. Virtual machine interpreter of pseudo-binary struggled with that, but compiled C worked just fine.

To accommodate minuscule (by modern standards) memories standard at the time and the slow speed of disk access, I nicked from Prime an outline of their paging algorithm, and built that into the kernel, together with the ability to report locate demands, the number of locate buffers being specifiable at compilation time. This mechanism is still present within the kernel, and is required to build DOS versions of games. The kernel also provided the option of reading all texts from the data file as required, with no paging mechanism, or to load the data file entirely into memory on startup. Initially there was no option for pre-loading all of the text into memory -- compilers of the time tended to choke on that.

Unix (Irix in fact) and later Linux became the default platform. Initially I built DOS/Windows executables by using djgpp under DOS. Later DOS and Windows builds diverged. The DOS version is now created using djgpp (run under wine), whereas the Windows version gets built using MinGW and packaged using InnoSetup. At some stage a Mac build was added as well.

As memory capacities grew and general machine performance improved, I first added the option to build the whole game as a single executable with no data file – all text begin stored in initialised arrays in the C source. Later this became the default. Thus things stood until I embarked on Adv770, which is when I discovered some unexpected upsides and downsides of the approach I'd taken.

I needed the game to be tested by others, but I knew from experience that it would be foolish to rely on testers to report all problems. Without knowing what is supposed to happen, it is not necessarily obvious whether things are going wrong or not. I simply had to have access to testers' log files. The only way this could be achieved was to run the game through a web interface – in the cloud, as one would say these days. Unfortunately, as the acdc/kernel implementation of A-code stood at the time, this simply was not possible. An A-code game was simply an executable, taking player commands from standard input and responding on standard output in a continuous loop.

So I sat about to implement a single-turn operation mode. In this mode, instead of waiting for play command, the game would automatically dump its current state to disk and exit. And when a new command arrived via a web interface, the game would be restarted and the saved state automatically loaded before processing the command. This would be easy enough with Dave Platt's original virtual machine approach, but my translate-compile-and-run implementation entailed that in an A-code game there could be only a single place where the player could be asked for input. Specifically, any use of the QUERY directive, used in Adv550 and Adv660 to handle yes/no queries, was simply out.

My eventual solution was to implement the context mechanism and replace any use of query with setting the special variable CONTEXT to a unique value, specifying the nature of the query, before saving the game state and exiting. Once a response arrived from the player, the game would restart and if the value of the restored CONTEXT indicated that a yes/no question had been asked, it would evaluate the answer in the same way as it would have done if QUERY had been used. This permitted cloud-based operation (initially CGI, later PHP). It also, quite unintentionally, enabled all A-code games (even Dave Platt's original code of Adv550) to have a persistent state. If a game was simply interrupted (or... erm.. crashed), it could be restarted from that point without being explicitly saved by the player.

Some internal changes had to be introduced to support this development, and so the major version number of the acdc and the A-code kernel got incremented from 10 to 11. Version 11 also brought in another innovation. All earlier versions assumed games being played in a terminal emulator, and assumed the display to be limited to 80 characters per line and 24 lines per screenful. Version 11 added the ability to run a game in its own window, initially by using the GLK library.

On the plus side, since there was no virtual machine to save the state of, and I'd avoided using the QUERY directive, it was pretty easy to maintain upward compatibility with testers' saved games, despite the game undergoing some major bug fixing.

Adv770 was finally released in 2003, but I continued tinkering with my A-code implementation anyway. 2008 brought version 12 of A-code, which removed the requirement of declaring game's entities before they could be used. This was achieved by changing the acdc translator to make two passes over game source instead of one &ndash the first pass collected information on all game entities being declared by the source and the second pass actually translated the source into C.

In changing to version 12 I also took the opportunity of ditching GLK because its Unix/Linux implementation of GLK had severe limitations (and used really ugly, hard to change fonts). Instead an A-code game executable could launch the default browser and then act as a very simple HTTP server, using the browser as the user interface.

In 2013 Brian Ball asked for changes which would enable him to port Adv770 to iOS. This was an interesting challenge, since iOS demanded to be in charge of the game's command loop. So the kernel was twisted yet again, to allow a "library mode". Fortunately, this was not too hard to do, since the game persistent mechanism introduced for the CGI/PHP operation could be adapted for the purpose. The main difference being in returning accumulated text to the calling routine, instead of displaying it to the player.

This change had a large unexpected payoff a year later, when I used emscripten to translate the C-code generated by acdc into JavaScript, so that it became possible to run the game entirely in an HTML 5 compliant browser.

Finally, in 2020 (doesn't time fly!) I got around to making my implementation of A-code to be entirely UTF8 compliant. While UTF8 encoding could always be used in game texts and object/place descriptions, the challenge was to permit UTF8 in entity names and hence in player vocabulary. That done, I also extended non-vocabulary names convention to vocabulary word declarations which frees A-code games from any dependence on the English language by allowing one to safely re-define the default command parsing words AND, THEN and AGAIN. So if you want to write IF games in Japanese or Russian, you can! :-)

And that's where things stand at the time of writing.


Back to the documentation index
To the Mipmip home page
Feel free to leave a comment!
Mike Arnautov (27 March 2023)