This page is meant to provide some basic suggestions and strategies for people who are starting out with reverse engineering old adventure games, and aren't sure how to do it. It mainly focuses on resources and tools for reversing DOS game executables, but much of the strategies discussed may apply equally to other systems and debugging tools. This is only intended as an overview; you'll still need to read other resources to learn 8086 assembly language, and learn how to use the various tools effectively.
IDA is one of the best disassemblers available. And luckily, the freeware version works with old DOS executables. Even more so, the current freeware version supports viewing disassemblies in graph mode, making it easier to see the overall flow of individual methods.
The DosBox Debugger is an invaluable tool for running old DOS games, to monitor how the program executes, and what values are generated by the executing code.
XVI32 Hex File Viewer
Although IDA has a built in hex viewer for the executable itself, the XVI32 tool is useful for viewing the contents of all the other files that come with a game. There are many different freeware hex editors available, so any other can be used just as easily.
Ralf Brown's Interrupt List
A nice reference for the operation of DOS interrupts. In 8086 assembly, apart from directly accessing ports, using interrupts is the primary means of accessing system functionality such as opening files, changing graphics modes, and many other things.
8086 Assembly Language
For those new to 8086 assembly language, you'll need a handy reference to learn the syntax. The Wikipedia is a good starting point, but you can also simply Google for an introduction as well.
"Reverse Engineering for Beginners" free book
This site has a free ebook that may be useful as a gentle introduction to reverse engineering techniques in general.
A pretty cool online tool that lets you paste in C code and shows you the compiled assembly under various different compilers. Useful if you're familiar with C, and what see what kinds of assembly are produced for various different code fragments.
FRIDA - Dynamic Instrumentation Framework
Nice tool for a easy writing and injecting hooks to game binaries. Useful for watching how code executes, when are internal functions called, for dumping structures from memory of target process and for changing data in memory on the fly. Works for newer binaries of games (32bit and windows xp)
Using the DosBox Debugger
It's up to the individual if you want to use a debugger when reverse engineering a program. Some prefer a more cerebral challenge of only figuring out code execution using a decompiler tool, whereas others may find using a debugger useful for figuring out what values are passed to functions. I would recommend using a debugger particularly when reversing a game for the purpose of adding ScummVM support. When you start implementing code to implement game functionality, once you've got portions of the game disassembled, it can be immensely useful for tracking down bugs. Particularly if you initially write your code with names that closely match the names you give the methods in the disassembly.
For debugging purposes, if the game is a DOS game, the DosBox Debugger is the best tool I've found for executing and debugging DOS programs. The default distribution of DosBox doesn't have it enabled, but you can either compile DosBox with it enabled, or download a previously compiled executable. See the DosBox Debugger Thread for more information.
One of the biggest initial steps when using the DosBox debugger is matching addresses in executable at run-time with your disassembly in IDA. This can be done either from the debugger to IDA, or from IDA to the debugger:
From the debugger to IDA
This is the easiest. If you break execution of the game at any point, you can simply use the Find Binary option in IDA to search for a sequence of bytes from the instructions shown in the DosBox Debugger disassembly area. Be careful to pick instructions that aren't far calls or jumps - such instructions are modified when a program loads depending where it loads in memory, so the IDA disassembly won't have the exact same bytes. If you do find a match, double check that the offset within the segment of the found match in IDA matches the offset of the instructions in the DosBox Debugger. If not, you may have found a false match, and should either search for the next occurrence, or specify extra bytes in your search until you find the correct match.
From IDA to the Debugger
If you have a point in the IDA disassembly and want to figure out what address it will be loaded in the DosBox Debugger, it's also not hard. This is presuming the game in question doesn't use Overlays. Overlays were a method developed when games and other applications became too big to fit into memory at once. In these cases, code for the program is often stored at the end of the executable, or in separate files, and loaded as needed into part of the memory, overwriting previously loaded code. In such situations, it becomes hard to pin down a specific section in memory a given segment will be loaded at, since it may shift around in memory over the course of the program running, as it gets loaded, overwritten, and loaded again repeatedly.
So long as the game doesn't use overlays, the following steps can be used:
- look at the IDA view to find out the current file offset at the bottom of the screen. You'll quickly find it if you try selecting different instructions, since it will keep changing. Now:
- Get the value from the beginning of the current code segment. This is just to make the calculations easier, since the start of the segment will have an instruction offset between 0h and 0Fh, which means it won't be messing with
our segment calculations
- Get the value from the beginning of the entire disassembly.
- Drop the last digit from both values, and get the difference between the two.
- For executables run in DosBox, add a value of '0138h'. For COM files, add a value of '0128h'.
This will give you the segment address of where the segment should be under DosBox. In either case, it's generally a good idea is to then rename the current segment in the IDA disassembly so that it includes the actual segment address of where it was loaded in DosBox.
For example, the first segment of executables is normally loaded at segment 0138h in memory, so you might rename the segment 'sg0138'. That way, if you later want to set a breakpoint in the DosBox Debugger for any instruction in the segment, you will immediately know what the segment is.
Using IDA Effectively
One of the best things to do when disassembling a game is to document everything. Particularly method parameters and structures.
Methods can be renamed using the general 'N' hotkey (as well as via the menus), and the 'Y' can be used to specify a C-like prototype for a method. This is particularly useful when some of the parameters for a method are passed using registers. By explicitly documenting what the method expects, it makes it easier to remember later on when you're reversing methods that call it. Standard methods where parameters are passed via the stack are easy, since IDA can automatically set up the function prototype for you. If a method does have parameters passed in registers, prototypes like the below can be used:
int __usercall sub_100FB<ax>(__int8 param1<al>, int param2<bx>)
In this case, the method takes an 8-bit parameter in the al register, and another 16-bit value in bx, then returns a result in ax
The other thing you'll need to learn to use IDA effectively is the use of structures. Irrespective of what language a game was originally written in, there will always be structures containing related information. It may be something as simple as a C-style struct, or could even be the fields of a class in C++.
When dealing with data, you'll frequently see cases like
mov bx, 30h
mov ax, [bx+2D00h]
In this case, an initial index in the ax register is multiplied by 30h (30 hexadecimal = 48 decimal). So from this we can determine that the given structure is 48 bytes in size, and can create a new structure accordingly. For smaller sized structures, you may want to create as many 2 byte word fields as needed to make up the correct size for the structure. For larger sizes, the easiest way is to simply declare an array of the needed structure size - 1, and follow it with a single byte field. You can then delete/undefine the array. The remaining byte will keep the structure at the correct size, and you can then later fill in the fields as you find references to them.
Secondarily, the value of '2D00h' indicates an offset in the data segment, representing the rough starting address of the first element of the given array in memory. Here we run into a minor problem. The offset of '2D00h' may not indicate the precise start of the array. If the code in question wanted to get the value at offset 8 in the structure, then the array may actually start 8 bytes earlier in memory, at address '2CF8h'.
In such cases, the only way to tell for sure is to start searching for immediate values in the program of values bytes backwards at a time, until you can't find any more values. For example, if you find references in the code of values '2CFFh', and '2CFEh', each with previous multiplications by 30h/48, but none for '2CFDh', '2CFDCh', or '2CFBh', then you can probably be confident that the array starts at offset 2CFDh.
Once that's determined, you can then create a dummy structure of the correct size, and convert the given address of 2CFDh to an instance of that structure type. Until you're more familiar with the range of values the original array index may be, it'll likely be easier to simply leave the defined array with the single index. Later on, you can always change the structure in memory to specify how many elements it has later on.
Remember that fields in structures can vary in size, so it's always possible you'll get the starting address wrong. In which case, you may have to later on correct the address of the structure in the data segment. This will affect any fields you figure out as well. In the above example, if you mistakenly presumed the array started at offset 2CFDh, then 2D00h would be thought to be a field at offset 3 in the structure (2CFDh + 3 = 2D00h, as per the above example code fragment). However, if the array structure really starts at 2CF8h, then the same field should be at offset 8 within the structure (2CF8h + 8 = 2D00h). So you need to rebuild the list of fields you'd figured out in the structure, since they'll all be at the wrong position. Overall, it's better when encountering an array to spend the extra time to ensure where it starts in memory so you don't need to fix offset problems later on.
One of the hardest things when starting work on a new disassembly is to figure out how to begin. The following are offered as suggestions of how to get started in the disassembly process.
One of the easiest places to start a disassembly is generally by identifying file accesses. Using IDA, you can, for example, do a text search for 'open', 'read', 'close', etc. to find occurrences of file opening. IDA provides standard comments for many operating system calls, so even in a new disassembly you should be able to locate such calls by their comment text. Likewise for file reading, writing, and closing. Normally, a program will encapsulate these calls into a method of it's own, so your first disassembly step can be in identifying the methods and naming them appropriately with names like 'File_open', 'File_read', and so on. Likewise, giving the passed parameters an appropriate name. In IDA, the 'Y' command can be used to set up an appropriate method signature for methods. By properly naming the method and it's parameters, this will help you in all the methods that call those methods.
For example, if a read method has a 'size' parameter and a 'buffer' parameter, then if a method that calls it passes '200' for the size, and a reference from a location on the stack, you can be confident that the stack entry can be called something like 'readBuffer', and use the '*' (array size) key when looking at the Stack View (Ctrl-K) to set the size of the array to 200 bytes.
You should hopefully then be able to start working on methods that call the file access functions and hopefully start decoding them. Some examples:
1. If the game consists of only a few large data files, the methods that call the open/read/close functions may a resource manager responsible for loading subsets of the file. In which case, the methods may likely load some kind of index into memory and then have a separate 'get resource' method that scans through the list for a resource with a given Id, resulting in a specific portion of the data file being read. In this case, you can identify all the methods with appropriate names like 'ResourceManager_init', 'ResourceManager_loadIndex', 'ResourceManager_getResource', and so on.
The DosBox debugger may prove useful when dealing with games using large resources. In DOS, Interrupt 21h is one of the primary system interrupts. Specific command Ids are passed in AH, and the other registers are set with values depending on which function is being called. For example, command 42h of Interrupt 21h is the command for seeking within a file.
Try using 'BPINT 21 42' to put a breakpoint on any calls to seek system function. By clearly identifying the 'Seek method', you can then step out of that routine to find what called it. Hopefully, you can then examine the logic in the disassembly used to generate the file offset to help you figure out how file offsets are generated for specific resources, and from that figure out how the resource's index works.
Remember that game resource managers not only typically merge multiple individual resources into one single bigger file, they frequently also compress them as well, to save space and prevent people from seeing textual resources when viewing the contents of the file. In such cases, if you can figure out the strategy used for extracting single resources, it may be worthwhile taking the time to code a standalone program to extract and, if necessary, decompress single resources into separate output files. That way, you can more easily look at individual resources that are used by the game without having to worry about manually locating them in the archive/resource file.
2. If the game consists of many different files, it's likely the game will be manually calling the open/read/close methods whenever it wants to access a particular file/resource.
In either case, figuring out the file access routines will give you an excellent start into figuring out the contents of the game; you can then move onto methods that call the resource manager get resource method, and start looking at what kind of resources are loaded, and from there start identifying methods that make use of those resources.
Another place to get started on the disassembly is the graphic draw routines, those responsible for copying raw pixels to the screen surface.
Graphic display was complicated in the early PC days by different modes for the different graphics cards writing to memory in different ways. In the Monochrome/Hercules mode, for example, 8 pixels are stored per 8-bit byte. In EGA, the addressing can be complicated by how the display is configured - the same areas of memory may be used to represent different parts of pixels - with the part of a pixel being updated depending on specific values sent to hardware ports. Finally, of them all, the most common 320x200x256 color mode is the easiest to deal with, with each pixel taking up a single byte.
For most of the graphics modes, you can look at them in a similar manner - as a block of data in memory starting at offset A000h:0. Only the number of bytes per line will vary, depending on what the graphics mode is. Assembly routines that deal with the graphics screen will typically have code to figure out screen offsets based on provided x and y parameters, so it will frequently be easy to identify the parameters and figure out how the screen offsets work. For example, in 320x200x256 MCGA mode, an offset on the screen will be calculated using the formula (y * 320) + x.
For finding the graphic routines you have two options:
The first is to entirely use IDA, and simply search for immediate values of 'A000h'. Since this is the area of memory that graphics are commonly displayed in, it can be a quick way to locate graphic routines.
The other alternative is to use the DosBox Debugger. It has a use command called 'bpm' that allow you to set a memory breakpoint, which then gets triggered if the given memory address changes. So you could do 'bpm A000:0' to set a breakpoint on the first byte of the screen memory (i.e. the top left hand corner of the screen). Then whichever routine modifies it first will trigger the breakpoint. Using the previously discussed techniques, you can find the same place in your IDA disassembly, and look into reversing that method first.
It will be likely that related functions will be next to each other, so once you've looked into the given identified function, you may also be able to review previous or following functions to see if they have identifiable graphic routines.
Data Segment strings
The strings in the data segment can be an excellent source for identifying the purposes of various methods. If you're very lucky, there may be error messages that contain the name of the function as part of the error. In which case, you can then find out what code references it, and name the methods appropriately.
Note: IDA is good, but it's not perfect., It's not always guaranteed to be able to figure out that a given value loaded into a register somewhere in the program is for a reference into the data segment. As such, if the cross-reference command doesn't give any references for a given string, try searching for an immediate value of the offset of the string. Chances are, any reference you find is likely pointing to the string. You can then use the 'O' command to change the operand from an immediate value to instead point to the offset in the data segment.
Even if any error messages don't contain method names, an error message can prove invaluable. For example, an error message like "Unable to initialise mouse" tells you that whatever method uses it is setting up the mouse for access, so could be given a name like 'initialise_mouse', or 'initialise_events'.
Likewise, have a look at the context of what needs to happen for the error message to be printed, since the message can give you insights into what is being done. A message like "No more free inventory slots" tells you that the code that references this error message is likely a routine for adding an item to the inventory (hence the error if no more slots are available). From there you might be able to identify the area of memory containing the inventory list, and then cross reference other methods that also access it - you could end up identifying a whole group of methods related to inventory manipulation.
Another possibility for data segment strings is to pick strings that may be descriptive enough to guess their purpose. For example, if a string has a name like 'FNT20', or 'UI.FNT', it's likely that it's a font file, containing the images for each character for use in displaying on the screen. In that case, any code that references it is likely to be passing it to a 'font_open' function, which loads up the font. If you can disassemble that method, you may be able to determine how the loaded font is stored in memory. From there, you can use the cross references function of IDA to find other methods that use the same memory address as where the font is stored, and which will likely give you the methods for actually displaying text on the screen.
From there, you may be able to go even further, and start figuring out methods that call the 'write string' function.. menus, hotspots, conversation handlers, and so forth.
Another method for identifying methods may be simply executing the program itself. When running the program in DosBox, you may find it useful to start stepping through the main procedure to see what happens from stepping over every method. IDA may be helpful in identifying the main procedure, but even if not, most programming languages have a series of method calls for setting up the initial application state,and then a single final call to the 'main' method.
With the main method identified, stepping over each method may produce interesting results. For example, stepping over a single method call may run all the code for showing the game's introduction sequence before returning control to the debugger. If this can be identified, you could name the method appropriately. You may then be able to gleam information from how the method is called and for the method itself:
For where the method is called, there may be conditional checks to see whether the method is called or not. For example, some games may have a stored settings file to flag whether the introduction has been shown, and not show it again after the first time the game is played..
In that case, a method call just prior to the method call to show the introduction may be for reading in the game settings and checking the flag for whether the introduction has been shown. Knowing this, you could name multiple methods for reading settings, then within it file opening and reading, and so on.
Within the method itself, you may likewise be able to figure out further specific details of how the method is implemented from the method calls it makes. For example, a 'play introduction' method may consist of calling the same method multiple times with different parameter values. These values may well be offsets within the data segment for resource or file names for specific animations to run. In which case, you now know the sub-method is an animation player, and name it accordingly. You could then start work on the animation player, figuring out how it loads it in data, and what method it uses to build up and display graphics on the screen.
Particularly for cases like that, identifying and naming the graphic/screen methods may be helpful, since you could work the disassembly from both the front end reading the animation, and from the low level drawing of the graphics of the animation.
Following are a few other examples of specific strategies I used to make initial progress in figuring out adventure games I've worked on in the past. They may help give you inspiration for your own work:
The font system
I frequently like to tackle the font system for a game early on, since it's usually fairly straight-forward to figure out, and it's a nice milestone for your own re-implementation of the game to be able to display visual text.
- I started by taking a screenshot of the game that showed some text, and used Paint to determine the screen co-ordinates of one of the pixels in a character that was displayed on-screen.
- I then calculated what the memory location of that pixel was in the screen. Presuming the fairly standard 320x200 8-bit graphics, the segment would be A000h, and the offset would be y * 320 + x.
- Using the DosBox Debugger to run the game up to a point just before the text was displayed, I set a memory breakpoint for the pixel, and then allowed the game to start to display the text.
- The point where the pixel set was identified as the routine for drawing a character. It's fairly standard for it to take in an x and y position, and the specified character to display. I gave it an appropriate name like 'font_write_char'.
- By examining how the character is used to lookup the pixel data to copy the screen, I was able to identify where the font was stored, and what structure it had.
- By looking at other places that referenced the same memory I was able to identify other font related methods, such as for loading the font, or determining the width of a given character. And then some of the callers of these methods could then be identified. For example, one caller called the method in a loop for every character of a string, so I was able to call it 'font_string_width'. Further, since that method was now known to take in printable strings, I was able to see what called it and identify string handling in other parts of the game.
- Similarly, I was able to identify a method that called font_write_char in a loop, and could name it 'font_write_string'. As with the identified font_string_width, I could then look at callers to font_write_string to identify how other parts of the system pass strings.
- Because the display of information is so important for adventure games, the font_write_string method becomes an excellent starting point for all sorts of further investigations. For example, suppose in your adventure game if you try to combine two arbitrary items together and it prints "That doesn't work". You can set a breakpoint in DosBox debugger in the font_write_string method at it's end, and when it's hit by trying to combine the items, trace backwards to find out which method called it.
- Based on a similar test, I was able to identify methods that worked with the inventory, and the lists in memory of all the available items as well as the list of items in the player's inventory. I then set up an appropriate structure for items. I didn't know all the fields they held initially, but some fields, such as a pointer to the textual name, were easy to figure out, and I could always come back to the structure later and add proper definition for other fields when I figured them out. I then went on to identify other methods that also access the inventory list and identified the methods that handle adding and removing items from the inventory. I was also able to see the method that gets the details of the inventory for displaying the inventory on-screen. This was extremely useful, since it used a standard "sprite draw" method to draw both the inventory area background as well as the glyphs for each item in the player's inventory. And with the sprite drawing method identified, there would be the potential for looking at all it's callers and identifying methods that handle drawing in-game screens and dialogs.
In fact, even other inventory methods like adding and removing proved to be important too, as I was able to check their callers, and identify them as methods that were part of a script engine. These methods were indexed by a table of all possible script command handler methods, so I was able to see which method was using that list to identify the main script execution method. By setting breakpoints in the method and doing various actions in-game, I was able to start figuring out what the various script methods did, and based on that, identifying and naming the methods that they called. A pretty good result from a process that started with a single pixel on the screen.
The Hotspot list
In adventure games, hotspots are areas of the screen that has an interactable item. Moving the mouse over the area causes a description of the hotspot. Using a breakpoint in the write string method, I was able to see what the caller was. Then examining the caller to find out how the description came to be passed, I was able to figure out the structure of how the list of hotspots were stored in memory. This allowed me to create a Hotspot structure, and set up the array of hotspots in memory. From there I was able to go in two direcitons. Firstly, by identifying other methods that access the same hotspot list, then finding out which one set up the values, I was able to locate the hotspot laoding code, which was part of the overall scene loading code. With scene loading identified, I could look at the other files being accessed, and the structures being loaded, and get further ideas of what information each scene contains.
The other direction I was able to go in was hotspot interaction. By setting breakpoints in other methods that accessed the hotspot list and then trying to interact with a hotspot, I was able to identify the general method that handles actually doing item interactions. In this case, it turned out to use other fields of the Hotspot record, and then load a script and call it to be executed. Since I'd already identified the script execution method, it made it easier to realize that script data was being loaded, since it was immediately passed on to be executed.
Scenes would be kind of boring if they didn't have any animation or action within them. As such adventure games tend to have a main scene drawing method to draw in the background and then add in any sprites for extra items within the scene. I was able to take an item that's obviously animating, and then set a memory breakpoint on a pixel within it just before the new scene was loaded, and see what changes it. The first change was clearing the screen. Next was rendering of the scene background, which in itself was a useful diversion to figure how the scene specified the background to load, and what format backgrounds are stored in. Then the next change came from rendering the sprite onto the scene. In this, I confirmed that the scene items were using the already identified sprite drawing method. So by tracing out of the sprite drawing method (by setting a breakpoint on it's 'retf' opcode and then running DosBox debugger until it it's hit, you can then single step out of the method), I was able to identify the code looping to draw items within the scene, and the overall method that draws the contents of the scene for each frame.
I was also able to start figuring out how the structures for the scene sprites were being stored in memory, with fields such as the position for where to draw the sprite, which set of sprites to use, and what frame number. The frame number particularly is useful.. sprites in games generically collect multiple "frames" in a single sprite, representing all the possible ways the given sprite can be drawn. By seeing what method was called to get a specific frame to pass to the sprite drawing code, I was able to locate the methods that handle managing sprites, and flesh out both it as well as other nearby methods in the disassembly that handled loading a sprite from files.
In fact, at this point, the only thing I was missing was loading in an appropriate palette to be able to implement the basics of loading a game scene, drawing meaningful sprites within the scene, and having hotspots work. Presuming any other adventure game you look into has similar kinds of structures, this could lead to you having a concrete start at reimplementing the game, and enough methods figured out in the disassembly, and pick any area of the game you want to work on.. at this point you'd have the basics of scene loading, graphics rendering and sprite display, scene interaction, and the basics of actually doing items. Maybe you could..
- figure out which sprite controls the player, and put memory breakpoints on the sprites position and/or frame. From that, you could identify methods that control moving the character, and from there investigate the pathfinding logic the game uses, leading to you being able to have your re-implemention's character walking around the screen
- Flesh out more of the script system. Maybe concentrate on some of the first actions done in the game, so your reimplementation can likewise do them.
- Look into the music/sound system. From strings in the data segment, or the filenames of specially loaded music drivers, you may be able to figure out where the sound code is loaded, and start looking into what commands the sound and/or music drivers have, and what sound formats are used.
There's lot of areas you can proceed to delve in. Hopefully the suggestions above will help get you over the initial barrier of not being sure how to start, and assist you getting started in your efforts.
Reverse engineering a game can be a rewarding experience, but plan to spend a lot of time working on it. Reverse engineering a game can take months, even years for really complicated games. As you gradually start figuring out sections of the original game, it may prove helpful to dive into creating the engine in a fork of the ScummVM code. Re-implementing things such as resource managers, and file access, it may help you figure out the purpose of other methods more easily if you can see what the code you implement does, and thus get a better idea of what kind of data will be passed to other methods you haven't yet figured out.
It'll also help notify the rest of us what games are being worked on, and ensure multiple people aren't working on the same game. Plus, it'll be cool for the rest of us to keep track of how you're going. :)