Let's dive into The Wizard's Castle (source code)!
That is the first line of the 80s microcomputer BASIC game The Wizard's Castle initially written for the Exidy Sorcerer platform.
It's a REMark statement, a comment in that particular language. 10 is the line number, if you're unfamiliar with languages that had such things.
But the interesting part was this "_(C2SLFF4 mess. A typo or garbage? No. It appears verbatim in the source code as originally published in the July, 1980 issue of Recreational Computing.
Note: I'm going to jump back and forth between decimal (which is BASIC-friendly) and hex (which is programmer-friendly). Hex will be prefixed with 0x, suffixed with h, or will obviously have digits A-F in it.
Note: Thanks to Josh and Chris, my hacking buddies, for doing tons of the emulator and research work on this. It took hours, but they were good hours. It would have taken me 10x as long without their help.
Here's more of the source in context, but with irrelevant pieces removed and some spaces added:
In BASIC, a : is a command separator. POKE writes a byte value to the specified memory location. PEEK reads from one. Sorcerer BASIC used signed 16-bit numbers, apparently, so address -2049 is 65536-2049 or 0xF7FF in hex. We'll revisit that address later.
Calling the pseudo-random number generator (PRNG) RND() with a negative number seeds it. Old PRNGs liked to be seeded with odd numbers, thus the 2*T+1 expression forcing it odd.
And the USR() function calls a machine code routine.
And line 80 is the first use of T after it is initialized from the PEEK().
The proximity of these things makes it seem like they're all related to seeding the PRNG. Sorcerer BASIC didn't have a RANDOMIZE command, so the only options of the day were:
Wizard's Castle didn't do the first two. So it must be the last one.
Could it be that the REM statement has machine code that happens to encode to ASCII text? The Sorcerer did use ASCII encoding.
But it seems nuts that you could get useful Z80 machine code just from ASCII characters. Right?
Nevertheless, this seemed like a reasonable-enough crazy assumption to run with and see what we could find.
The USR() call is interesting—this calls a machine code function, but the details are system-dependent. But luckily the Internet has tons of technical manuals in various places and we were able to figure it out.
Address 259 has a JP instruction (unconditional jump) to a 16-bit little-endian absolute address.
And the POKEs were for addresses 260 and 261. This was the trampoline for the USR() function. You POKE the address of the start of your machine code into addresses 260 and 261, and then run USR() to call it.
Reversing them for little-endian, this means we're looking at address (1 << 8) | 218 which is 474. When we run USR(0), it jumps to the machine code at address 474 that should end in a RET instruction.
The argument for USR() (the 0) is stored as a 4-byte float elsewhere in RAM so the machine code can look at it if it needs to. It turns out not to be relevant for this use case. Also we're assigning the return value from USR() into T, but it's unclear to me what this return value is. However, it is irrelevant as we immediately overwrite T with the next assignment.
When you enter a line of Sorcerer BASIC, the interpreter tokenizes it and replaces commands like PRINT with single byte values:
And then it is stored like a linked list in memory where each row is a "node".
And the start of this linked list in RAM is at address 469 on the Sorcerer.
Let's see the mystery line again:
That means the node for that line consists of the following addresses and bytes:
and 474 is where the USR() function gets us! It is literally calling the text of the REM statement as Z80 machine code!
Could it be true? Let's try it out!
Here are the hex values for the ASCII characters:
There's the 0x00 terminator at the end, but that's just a NOP in Z80, so we'll ignore it.
Now, I don't want to get into the details of Z80 assembly here, but take my word for it that this code makes no sense. The addresses don't seem to point to anything in particular, who knows what HL has in it at the beginning, B is never read, the duplicated LD B is pointless, and there's no RET to get us back to BASIC.
It's garbage. And running it in an emulator results in Weird Things happening (like soft resets), as expected.
This was a dead-end for the time being. We speculated that maybe it was a printing error, that the printer didn't have the glyphs for the actual bytes, so it substituted other ones. But that wasn't super-satisfying.
Let's switch gears, then, and continue on to the PEEK:
What's at address -2049? Unsigning this makes it equivalent to 0xF7FF. According to the docs, this is the last byte of memory-mapped screen text, representing the current character in the lower right of the screen.
We then immediately use this value to seed the PRNG:
I'm going to use my magical powers now to look at one of your open terminal windows and read the character in the lower right of the screen. It's fuzzy, but if I concentrate… it's… yes, it's a space character, isn't it!
Now, the space is 32, so seeding the PRNG with -(2*32+1) every game would make for bad replayability in this randomly-generated dungeon, so the USR() function must have something to do with it. Like it puts something on the screen at address 0xF7FF and then the PEEK() reads it. The screen is cleared not-long thereafter, so the player might not notice the blip of a character temporarily placed in the lower right.
The source code in its original dot-matrix glory
So we had strong suspicions that all this was part of the PRNG seed process. If only there were some way to prove it!
It was at this point that Josh noticed this line in the magazine where the program was published: "The first remark is a machine language routine to simulate the RANDOM function."
Sheesh. That'll teach me not to read things.
In BASIC, RND(), not RANDOM, is the random function, used with a negative argument to seed the PRNG, and a positive argument to get the next random number, which was historically a random floating point result between 0 and 1.
So what the author probably meant was RANDOMIZE which was popular in Microsoft BASIC for setting the random number seed to a specific value or prompting the user to enter one.
In short, it was supposed to be doing what we thought, but the machine code didn't seem to be doing that at all.
Josh and I had been running Sorcerer emulators in MAME (thanks to Josh for sharing the ROMs with me to get going) and Chris had another Java-based one that he'd gotten working.
In MAME, we'd entered the REM statement and looked in memory, and it had the nonsensical bytes and machine code as disassembled, above. And it didn't run properly.
But Chris found a tape image of the game and loaded that up in his emulator. And the first two lines of source looked like this:
WOW! Another(?) author has annotated the source with the hex of the machine code! Not only does it end in 0xC9 which is Z80 RET, but it has that lower-right screen corner address 0xF7FF right there!
How on Earth does the REM text map to that? Things like 0xED are clearly non-ASCII. But wait, some of the characters are, and their positions match what we see in the REM!
What actual values are there? Chris loaded the program and looked:
You can see the 0x00 null terminator down there. And a hidden space at the end of the line. And 237 is 0xED, and 95 is 0x5F… it matches the line 15 comment!
Let's get ready to disassembbbllllllllllleeee! I'll convert the numbers to hex and go for it.
Later, I'll find out that the part after the RET is disassembled incorrectly and that my annotation of which REM characters mapped to which hex values was wrong, but who cares for now! Everything is over at the RET anyway.
And the code is exactly what we were looking for! Let's look at the first assembly instructions:
What's it do? The R register is interesting on the Z80 in that it is incremented every instruction fetch. Maybe? Depends on which resource you ask. In any case, it is incremented very, very frequently relative to human timescales. And the Sorcerer busy-waits for input, so when you finally run the game, it effectively has a random-ish value in it.
But seeding many lesser PRNGs with 0 is a bad idea (they tend to just produce 0 after that), so the code retries if it happens to get a 0 in R. But if it is non-zero, it stores that value in address 0xF7FF, the lower right corner of the screen! And then the PEEK() picks it up and uses it to seed the PRNG!
The R register only increments the low 7 bits, though, meaning it can only be 128 different values. And we discard 0, so really this means on the Sorcerer, there were only 127 different randomly-generated dungeons that you could play. Bummer!
So the glyphs we see simply aren't ASCII. Our first disassembly was doomed to fail because the REM violates the assumption that all listing text was ASCII.
For verification, I made a new program that was just a REM followed by a bunch of spaces (to give me room to work before the null terminator). Then I POKEd in the right values and got a listing.
It worked! Those values gave me the "_( glyphs that were in the original source listing!
The whole idea in the 80s with magazines like this was that you'd get it in the mail or from the newsstand and you'd painstakingly spend hours typing them in and getting them right.
It wasn't easy. Here's another excerpt from The Wizard's Castle:
Just pure codevomit. But those of us who typed this stuff in were pretty practiced at getting it right. So when we saw something like this:
you can bet we typed it in exactly.
But now we know this wouldn't have worked if you assumed those were ASCII-encoded glyphs.
Perhaps there was some lore well-known to Sorcerer programmers where they knew the magic incantations to make it go.
Or maybe he just wrote this to get some space to work:
And then manually POKEd the machine code in like I did, above, to get this:
And then never mentioned how to do it, forgot about it in the excitement of getting it published, and out the door it went.
But what is the deal with that F4 thing? It's really bugging me.
An F-4 Phantom. It's an aviation/programming pun. I hereby indemnify myself against lawsuits.
Chris's version had some extra F4s:
And we did a memory dump of Chris's version and got the following:
Notice anything weird? About the F4s? I didn't either. There are three of them in the REM, but only two of them in the memory dump (the 70,52 pairs)!
And that's not all—where's the L? Something's wrong. Earlier, I'd tried a sanity check by POKEing in the first three characters by hand, but let's try the others.
I'm going to recreate the entire machine code here and see what pops up.
Then I run that, and list it. I get:
Wait! My Xs shifted right by two characters! What's up with that? You can't just insert characters like that. It's like there are two extra characters in the output. I poked eight values, but there are 10 characters printed before the Xs!
Let's dump RAM and see what's there. I'm going to annotate the output, but—spoilers—my annotation is wrong:
No Fs or 4s. It goes directly from my last 201 (in the DATA) straight to a bunch of Xs (the 88s). So why do I see them in the listing?
Let's get specific. I'm going to manually plug in the 255, the 247, and the 201 in the REM and see what we get out of it.
Okay, it's the S, as expected.
Whoa—what? LF? Two characters? And suspiciously like "linefeed", but who knows. But that matches the REM!
And there's the F4. So the last two bytes are being printed as LFF4. That's where the extra two characters are coming from in the output.
So fixing my memory dump annotation, it would look like this:
And now that's matching what we're seeing in the listing:
But that's not all. The bytes with values 128 and just over actually map to BASIC keywords! Check this out:
When BASIC is printing its lines, we speculate that if the high bit is set, it looks up the symbol name and prints that instead. The weird symbol names we get up higher (random-seeming) might be (again, we speculate) mapping into garbage past the end of that table.
But wait just a second. In Chris's memory dump, there were bona fide ASCII Fs and 4s in there
Why were they added when they have no effect on the machine code? I think we'll just have to let that one go as being lost to the mists of time. Or maybe you can try gazing into an orb until you find the answer. Just don't overdo it.
The only point of all this, since I basically knew up front it was seeding the PRNG, was to feed hacker curiosity. What a luxury!
Would this work on other platforms, I wonder? Or would they just freak out? Someone else can give it a try on the Commodore 64 or something.
If you want to know more about The Wizard's Castle and even play this bit of history, I have a collection of documents and information on GitHub.
Source: Hacker News — This article was automatically imported from the source. Read full article at original source →
