2011
06.10

Part 1 of this article talked about how I found the bytecode files and the interpreter assembly code. Let's go further and try to understand how these bytecode files are structured.

We now have 188MB of JSON describing the CPU state when an instruction tries to access either the bytecode file in memory or the interpreter state. However, it is not readable at all:


{'type':'r', 'size':4, 'addr':'016E1964', 'val': '00007D40', 'pc':'80091DB8'
{'type':'r', 'size':4, 'addr':'016E00F4', 'val': '11000000', 'pc':'80091DBC'
{'type':'w', 'size':4, 'addr':'016E1964', 'val': '00007D44', 'pc':'80091DC8'
{'type':'r', 'size':4, 'addr':'016E0544', 'val': '00000000', 'pc':'80092720'
{'type':'r', 'size':4, 'addr':'016E053C', 'val': '816D83B4', 'pc':'80092724'
{'type':'r', 'size':4, 'addr':'016E1964', 'val': '00007D44', 'pc':'80092730'
{'type':'r', 'size':4, 'addr':'016E00F8', 'val': '00000000', 'pc':'80092734'
{'type':'w', 'size':4, 'addr':'016E1964', 'val': '00007D48', 'pc':'8009273C'

We need to enhance this by using the informations we already have on the interpreter behavior (where PC is stored, where is the instruction dispatcher, etc.). I wrote a very simple script called filter.py, which iterates on each line, loads the JSON object as a Python dict, applies a list of filters to it and prints the filtered line. Here is an example of a filter which detects lines in the dump from the instruction dispatcher:

(lambda d: d['pc'] == '80091DBC',
 'ReadInstr: %(val)s at pc=%(r3)s (@ %(pc)s)',
 'blue'),

It's a tuple with a predicate function, a format string which will be formatted using the JSON object itself, and a color which is translated to the equivalent ANSI color code (for blue, it is \e[34m). We can also write filters for instructions trying to access the script bytecode, and instructions manipulating the Program Counter (PC):

(lambda d: (int(d['addr'], 16) - 0x016E1964) % 5152 == 0
           and int(d['addr'], 16) >= 0x016E195C
           and d['type'] == 'r',
 '  GetPC: %(val)s at addr=%(addr)s (@ %(pc)s)',
 'green'),
(lambda d: (int(d['addr'], 16) - 0x016E1964) % 5152 == 0
           and int(d['addr'], 16) >= 0x016E195C
           and d['type'] == 'w',
 '  SetPC: %(val)s at addr=%(addr)s (@ %(pc)s)',
 'red'),

(lambda d: 0 <= int(d['off'], 16) < 0x8191,
 'SoAccess: type=%(type)s val=%(val)s at off=%(off)s (@ %(pc)s)',
 'yellow'),

Now that we know where the PC is stored, a first step would we to locate the control flow handling instructions. I took all of the ReadInstr lines from the dump and analyzed the PC value to see which instruction was doing jumps, i.e. after which instruction is the PC at least 0x10 more or less than its previous value. I won't paste the script here (it's easy to code but a bit long), but it was of a great use, finding only four instructions modifying the control flow. Opcodes starting by 05, 06, 08 and 09 all modified at least at one time in the dump the value of PC. Looking at the dump, 05 and 06 seems to do a lot of stuff, storing and loading from addresses we don't know about yet, but analyzing the PC after and before those opcodes, we can determine easily enough that they are the usual CALL/RET instructions. For example:

Opcode 05 at 7D48, jumping to 742C
  Opcode 05 at 7434, jumping to 6FA0
    Opcode 05 at 6FB8, jumping to 6D8C
    Opcode 06 at 6D88, jumping to 6FC0
  Opcode 06 at 6F9C, jumping to 743C

See how the jump address from opcode 06 are always just after an opcode 05? Looking at the dump a bit closer, we can also see that op05 writes its next instruction PC somewhere in the interpreter state (let's now call this somewhere "the stack"), and op06 reads it from exactly the same place! Let's look at opcode 08 now in the memory accesses dump:

ReadInstr: 08000000 at pc=00006E24 (@ 80091DBC)
  SetPC: 00006E28 at addr=016E1964 (@ 80091DC8)
{'type':'r', 'size':4, 'addr':'016E0544', 'val': '00000000', 'pc':'800923EC'
{'type':'r', 'size':4, 'addr':'016E053C', 'val': '816D83B4', 'pc':'800923F0'
  GetPC: 00006E28 at addr=016E1964 (@ 800923FC)
SoAccess: type=r val=00006D80 at off=00006E48 (@ 80092400)
  SetPC: 00006E2C at addr=016E1964 (@ 80092408)
{'type':'r', 'size':4, 'addr':'016E0544', 'val': '00000000', 'pc':'8009240C'
  SetPC: 00006D80 at addr=016E1964 (@ 80092418)

So... it reads instruction 8, adds 4 to the PC, reads a word from the bytecode, adds 4 to the PC, then sets the PC to the word read in the bytecode. In other words, after executing 08000000 12345678, PC=12345678. That's an absolute jump, which seems unconditional: every time an opcode 08 is encountered in the dump, PC is modified. That means opcode 09 is most likely a conditional jump: they are almost always used to implement loops and if statements. Two parts of the dump related to opcode 09 seems to confirm that:

ReadInstr: 09000000 at pc=00000500 (@ 80091DBC)
  SetPC: 00000504 at addr=016E1964 (@ 80091DC8)
{'type':'r', 'size':4, 'addr':'016E0544', 'val': '00000000', 'pc':'80092420'
{'type':'r', 'size':4, 'addr':'016E053C', 'val': '816D83B4', 'pc':'80092424'
  GetPC: 00000504 at addr=016E1964 (@ 80092430)
SoAccess: type=r val=00000564 at off=00000524 (@ 80092434)
  SetPC: 00000508 at addr=016E1964 (@ 8009243C)
{'type':'r', 'size':4, 'addr':'016E0544', 'val': '00000000', 'pc':'80092440'
{'type':'r', 'size':4, 'addr':'016E115C', 'val': '00000000', 'pc':'8009244C'
{'type':'r', 'size':4, 'addr':'016E055C', 'val': '00000000', 'pc':'80092458'
  SetPC: 00000564 at addr=016E1964 (@ 80092464)
{'type':'r', 'size':1, 'addr':'016E0558', 'val': '00000000', 'pc':'80091CFC'
{'type':'r', 'size':4, 'addr':'016E0544', 'val': '00000000', 'pc':'80091D38'
{'type':'r', 'size':4, 'addr':'016E1970', 'val': '00000011', 'pc':'80091D44'
{'type':'r', 'size':4, 'addr':'016E0544', 'val': '00000000', 'pc':'80091D78'
{'type':'r', 'size':4, 'addr':'016E1970', 'val': '00000011', 'pc':'80091D84'
{'type':'r', 'size':4, 'addr':'016E0544', 'val': '00000000', 'pc':'80091DA8'
{'type':'r', 'size':4, 'addr':'016E053C', 'val': '816D83B4', 'pc':'80091DAC'

In this case, we jumped to 0564 with the opcode 09000000 00000564. However, another part of the dump shows us that opcode 09 does not always jump:

ReadInstr: 09000000 at pc=000065A4 (@ 80091DBC)
  SetPC: 000065A8 at addr=016E1964 (@ 80091DC8)
{'type':'r', 'size':4, 'addr':'016E0544', 'val': '00000000', 'pc':'80092420'
{'type':'r', 'size':4, 'addr':'016E053C', 'val': '816D83B4', 'pc':'80092424'
  GetPC: 000065A8 at addr=016E1964 (@ 80092430)
SoAccess: type=r val=000065EC at off=000065C8 (@ 80092434)
  SetPC: 000065AC at addr=016E1964 (@ 8009243C)
{'type':'r', 'size':4, 'addr':'016E0544', 'val': '00000000', 'pc':'80092440'
{'type':'r', 'size':4, 'addr':'016E115C', 'val': '00000000', 'pc':'8009244C'
{'type':'r', 'size':4, 'addr':'016E055C', 'val': '00000001', 'pc':'80092458'
{'type':'r', 'size':1, 'addr':'016E0558', 'val': '00000000', 'pc':'80091CFC'
{'type':'r', 'size':4, 'addr':'016E0544', 'val': '00000000', 'pc':'80091D38'
{'type':'r', 'size':4, 'addr':'016E1970', 'val': '00000011', 'pc':'80091D44'
{'type':'r', 'size':4, 'addr':'016E0544', 'val': '00000000', 'pc':'80091D78'
{'type':'r', 'size':4, 'addr':'016E1970', 'val': '00000011', 'pc':'80091D84'
{'type':'r', 'size':4, 'addr':'016E0544', 'val': '00000000', 'pc':'80091DA8'
{'type':'r', 'size':4, 'addr':'016E053C', 'val': '816D83B4', 'pc':'80091DAC'

See the missing SetPC? We can safely assume we know most of the behavior of the four control flow handling instructions: 05=CALL, 06=RET, 07=JUMP and 08=CJUMP. But now, we also know that there is some kind of stack used for CALL/RET. Let's add some new filters to our script in order to detect easily instructions manipulating this stack:

(lambda d: (int(d['addr'], 16) - 0x016E1160) % 5152 == 0
           and int(d['addr'], 16) >= (0x016E195C - 5152)
           and d['type'] == 'r',
 '  GetStackAddr: %(val)s at addr=%(addr)s (@ %(pc)s)',
 'green'),
(lambda d: (int(d['addr'], 16) - 0x016E1160) % 5152 == 0
           and int(d['addr'], 16) >= (0x016E195C - 5152)
           and d['type'] == 'w',
 '  SetStackAddr: %(val)s at addr=%(addr)s (@ %(pc)s)',
 'red'),

(lambda d: (int(d['addr'], 16) - 0x016E1968) % 5152 == 0
           and int(d['addr'], 16) >= 0x016E195C
           and d['type'] == 'r',
 '  GetStackTop: %(val)s at addr=%(addr)s (@ %(pc)s)',
 'green'),
(lambda d: (int(d['addr'], 16) - 0x016E1968) % 5152 == 0
           and int(d['addr'], 16) >= 0x016E195C
           and d['type'] == 'w',
 '  SetStackTop: %(val)s at addr=%(addr)s (@ %(pc)s)',
 'red'),

(lambda d: (int(d['addr'], 16) - 0x016E1164) % 5152 <= 0x800
           and int(d['addr'], 16) >= (0x016E1164)
           and d['type'] == 'r',
 '  GetStack: %(val)s at stack off %(soff)s (addr=%(addr)s @ %(pc)s)',
 'yellow'),
(lambda d: (int(d['addr'], 16) - 0x016E1164) % 5152 <= 0x800
           and int(d['addr'], 16) >= (0x016E1164)
           and d['type'] == 'w',
 '  SetStack: %(val)s at stack off %(soff)s (addr=%(addr)s @ %(pc)s)',
 'cyan'),

Let's re-run our filtering script on the dump and find instructions that access the stack! Opcode 11 seems to be quite small and modifies the stack top, let's look at it more closely:

ReadInstr: 11000000 at pc=0000000C (@ 80091DBC)
  SetPC: 00000010 at addr=016E2D84 (@ 80091DC8)
{'type':'r', 'size':4, 'addr':'016E0544', 'val': '00000001', 'pc':'80092720'
{'type':'r', 'size':4, 'addr':'016E053C', 'val': '816D83B4', 'pc':'80092724'
  GetPC: 00000010 at addr=016E2D84 (@ 80092730)
SoAccess: type=r val=00000008 at off=00000030 (@ 80092734)
  SetPC: 00000014 at addr=016E2D84 (@ 8009273C)
{'type':'r', 'size':4, 'addr':'016E0544', 'val': '00000001', 'pc':'80092740'
  GetStackTop: 000007B8 at addr=016E2D88 (@ 8009274C)
{'type':'r', 'size':4, 'addr':'016E0544', 'val': '00000001', 'pc':'80092764'
  GetStackTop: 000007B8 at addr=016E2D88 (@ 80092770)
  SetStackTop: 000007B0 at addr=016E2D88 (@ 80092778)

ReadInstr: 11000000 at pc=00000158 (@ 80091DBC)
  SetPC: 0000015C at addr=016E1964 (@ 80091DC8)
{'type':'r', 'size':4, 'addr':'016E0544', 'val': '00000000', 'pc':'80092720'
{'type':'r', 'size':4, 'addr':'016E053C', 'val': '816D83B4', 'pc':'80092724'
  GetPC: 0000015C at addr=016E1964 (@ 80092730)
SoAccess: type=r val=00000000 at off=0000017C (@ 80092734)
  SetPC: 00000160 at addr=016E1964 (@ 8009273C)
{'type':'r', 'size':4, 'addr':'016E0544', 'val': '00000000', 'pc':'80092740'
  GetStackTop: 00000794 at addr=016E1968 (@ 8009274C)
{'type':'r', 'size':4, 'addr':'016E0544', 'val': '00000000', 'pc':'80092764'
  GetStackTop: 00000794 at addr=016E1968 (@ 80092770)
  SetStackTop: 00000794 at addr=016E1968 (@ 80092778)

Opcode 11 seems to read an argument just after the bytecode in the file and substracts it from the stack top. For example, 11000000 00000010 removes 0x10 from the stack top. Most of the time this is done to reserve space for local variables (for example, on x86 you do sub esp, 0x10), so let's call this instruction RESERVE. Opcode 10 seems to do almost the same thing but adds to the stack top instead of substracting, so let's call it UNRESERVE.

Another opcode, 0E, seems to write values to the stack and substracts 4 from the stack top each time it does that. In other terms, it pushes a value to the stack. But where does this value come from?

In the next part of this article, I'll talk about the scratchpad, where the temporary variables are stored for computations, and how variables are represented in this interpreter. Stay tuned!

5 comments so far

Add Your Comment
  1. Wonderful written 😀 I lack the understanding of all this but it was a nice read anyway.

    Whats with the other file you talked about in the first part “ar.dat”? I know it contains the graphics for the skits but thats all I found out.

    • It’s a very simple archive format which is used to store the graphics data for skits, but also 3D models, animations, etc. for maps and cinematic scenes.

      I wrote a very simple program to extract those files: http://code.delroth.net/tos2-tools/src/1417154b625b/containers/ardat.py

      The format is the following (in a C-like language):

      struct header {
      uint32_t nfiles;
      uint32_t data_offset[nfiles];
      };

      Each file ends where the next starts.

      • I was not able to get this extractor to run… but with the information you gave me I was able to write a quickbms-script that extracts the ardat.

        #written by RetroHelix
        #All the real work was done by delroth. I only used his extractor to make this.
        #Tales of Symphonia ar.dat extraction

        endian big
        get nfiles long
        for i = 1 to nfiles
        get fileoffset long
        savepos currentpos
        get filesize long
        if filesize > 0
        math filesize -= fileoffset
        log i fileoffset filesize
        else
        get realfsize asize
        math realfsize -= fileoffset
        log i fileoffset realfsize
        endif
        goto currentpos
        next i

        It seems to work since I was able to get some tpls out of the file. But I wonder how the game knows where it have to look for the files. I mean the ardat contains another file of the same kind and again there are no names (no header).

        • If you look at the ar.dat for skit FC06_012 (just an example, all the skits do the same), the first file is another ar.dat archive, and the second file is a “copy” of the script (there are probably differences between this script and the internationalized one, this may be an artefact from the past 🙂 ).

          The first file of the second ar.dat archive contains a sequence of numbers. For FC06_12, it’s “40D 40E 410 414 414 417 41A 41C 41D 41E 472 473 476 479”. Note that there are as much of these numbers as files in the archive if you remove the one with the numbers.

          Another thing is that if you look at the syscalls executed by the skit, these numbers are kind of referenced when the characters of the skit are declared. For example, in FC06_012, when Marta is initialized, the number 40D is passed as an argument to the syscall which creates the character. 40D is the first number in the “index” file, and if we look at the first non-index file in the ARDAT, it is indeed a TPL with Marta.

          I’ve not completely reversed all of this yet but this “index” file seems to contain the informations the game uses to locate files in the ARDAT archive.

  2. I read about your hiatus from wiihacking some weeks (months?) ago on your twittersite and now you are undbubing a game and pctos seems to get updated again. Thats nice to see 🙂

    Ill try out Baten Kaitos after having enough Arc Rise Fantasia (it has a bad dub too but there are undubs floating around the net. mostly not fully functional but almost). Will you write something on the blog about the undubing process you are going trough?

*