Preface

The article tries to document the technical aspects of M.A.X. that could either be interesting for an enthusiast or could be relevant for an engineer.
The information found herein is not guaranteed to be complete or technically accurate.

Compiler Toolchain

M.A.X. v1.04 was built using the Watcom C/C++ 10.5 compiler. The original M.A.X. runtime is a 16/32 bit mixed linear executable (LE) that is bound to a 32-bit DOS extender stub.
The Watcom compiler was shipped with the DOS/4GW 32-bit DOS extender. The game was shipped with DOS/4GW 1.97.

The compiler supports various memory models from which M.A.X. used a Mixed 16/32-bit flat/small model.

C/C++ Runtime Libraries Floating-Point Libraries (80x87)
clib3r.lib , plib3r.lib math387r.lib , emu387.lib

The 3r suffix in the library names means that the compiler generates 386 instructions based on the 386 instruction timings and that the compiler uses the Watcom register calling convention [1][2].

In general the compiled code inherits the following characteristics:

  • function arguments are passed in registers EAX, EDX, EBX and ECX as long as they fit and then on the stack.
  • 48 bit __far pointers are passed 64 bit aligned in two registers or stack variables.
  • all registers except EAX are preserved across function calls.
  • functions with external linkage are suffixed with an underscore on machine-code level. E.g. main_.
  • variables with external linkage are prefixed with an underscore on machine-code level. E.g. __STACKLOW.

Libraries Used

M.A.X. is built upon several static link libraries. Out of the 5706 subroutines 1644 (28.81%) are coming from these libraries. Of course its also true that many of the library functions are not even used by the game.

Library Name Description
Watcom C 32-bit runtime (clib3r.lib) Implements standard C library functions like fopen, memcpy, etc. At the time the Watcom C/C++ compiler supported several platforms like DOS, Linux, Netware, Windows and in many cases the compiler supported various flavors of the C standard like POSIX, ANSI and ISO. Many of the related functions are not portable.
Watcom C++ 32-bit runtime (plib3r.lib) Implements standard C++ support. The library contains services that facilitate handling of C++ classes, global constructor and destructor lists, vtables, class inheritance and similar C++ “stuff”.
M.A.X. calls 32 global constructors from which 12 are compiler or system related. E.g. there is a global constructor called __verify_pentium_fdiv_bug(). Interesting… or not.
Watcom floating-point libraries (math387r.lib & emu387r.lib) M.A.X. is built with hardware floating point support, but if no x87 [3] hardware is detected on the PC at runtime it is able to fall back to software emulation.
GNW GNW is a user interface and OS-abstraction library designed and programmed by Timothy Cain [4]. Alternative versions of the library were used within Fallout 1, Fallout 2, Mapper 2 (Official Fallout 2 editor), Star Trek: Starfleet Academy, Atomic Bomberman and more.
GNW implements not only a windowing system, but memory and file system abstraction layers and more. It also provides a streamlined interface for debugging and a human input device recorder and playback service for arcade like attract mode. This is not the way M.A.X. realizes the demo gameplay in the main menu after a 60 seconds idle timeout though.
HMI S.O.S. v3.x HMI Sound Operating System [5] is a sound card detection and sound card abstraction layer for DOS and many other platforms. It supports digital and MIDI playback, mixing more than 20 sound channels, streaming a music sample directly from HDD to conserve RAM space and many-many more. M.A.X., Descent 1 and 2, Blood & Magic and the MVE library rely on the v3.x branch of the library. M.A.X. does not support MIDI music. The music in M.A.X. is too awesome to be in mere MIDI format. Thanks Mr. Luzietti [6]! M.A.X. wraps the SOS API layer into a game specific C++ Sound Manager component called soundmgr.
MVE Interplay’s own video player library [7]. Alternative versions of the library were used within many Interplay legends like Redneck Rampage Rides Again, Descent II, Descent II setup tool, Fallout, etc. The library integrates the HMI SOS library for sound playback and the VBE library for video display.
LZSS LZSS is used within the database manager component of GNW, but M.A.X does not rely on it.
VBE VBE is a VESA BIOS Extension interface wrapper layer to the video display. The VBE library is not used by the game itself. Only the MVE library relies on the VBE library functions for video playback. The game itself uses GNW for video display.

Watcom v10.5 C++ 32 bit Constructs

A bit of history

The first ISO C++ standard was published in September, 1998. ANSI standardization was proposed in 1989 and the first organizational meeting of the ANSI C++ committee took place already in December, 1989. The language specification continuously evolved during the time frame.

A very good publication about the early history of C++ can be found in [8].

Watcom implemented the full AT&T C++ 3.0 specification with support for templates and exceptions in 1993 within Watcom C/C++32 v9.5. Watcom C/C++32 v10.5 was released two years later in 1995. At the time namespaces, or runtime type information (RTTI) were just concepts or even less.

The following sections attempt to document how the Watcom C/C++ compiler organizes data related to C++ constructs in 16/32 bit mixed linear executables (LE) targeting MS-DOS hosts.
Most available publications about reversing C++ applications focus on state of the art compilers and related methods or data organization models. Most of those techniques or models simply do not apply to executables generated by the original Watcom C/C++ compilers and even the programming language was different.

The C++ user population doubled every 7.5 months or so between 1979 and 1991. The estimated number of C++ users was 400.000 in 1991. Why was C++ so popular?

One pretty obvious reason even for simple applications is semi automated resource deallocation or file handling via destructors. E.g. to load a character font using C, GNW opens a file, dynamically allocates a buffer for meta data, dynamically allocates further buffers for each individual character’s data, reads the file and fills all the buffers, and if any step fails the function enters dedicated else paths to deallocate all the resources necessary in a reverse order and finally it closes the file handle if necessary. All this is implemented by hand and every error condition has its own list of buffer deallocation and file handle operations depending on the already allocated and opened resources. This leads to bloated code with lots of conditional branches and repeated code snippets with minor differences only. Hard to read, error prone, difficult to maintain. On the other hand the generated code is efficient thanks to optimizing compilers that eliminate most of the duplicated code and many of the branches via various tricks and there are state of the art processes to eliminate or mitigate the risks of inroducing coding errors.
The same can be achieved in C++ in a much cleaner way. But do not be fooled, this does not mean that C++ is memory or file system safe by design, quite the opposite.

Used resources

  • The binary and source code releases of the Open Watcom compiler, and the last professional version with its source code, could be downloaded from [9].
  • The older professional versions are required as well to be able to perform comparisons, analysis on libraries and such.
  • The latest Open Watcom release for the favoured host OS is required for tools like wdis, wlib, wdump.

Analyzed code generator:
  WATCOM C/C++32 Compile and Link Utility Version 10.5
  wcl386.exe (1995-07-11)
  MD5 hash: 2af860b0f1e431852f4989f4226f3938 *WCL386.EXE)

Identifying the compiler revision used

The simplest way to narrow down the list of potential compiler versions is the copyright notice which is embedded into the startup code: WATCOM C/C++32 Run-Time system. (c) Copyright by WATCOM International Corp. 1988-1995. All rights reserved.. The string is very talkative. C++ support was introduced in 1993 within v9.5. 32 indicates that the 32 bit runtime library is in use. The copyright date code 1988-1995 clearly indicates that the compiler is not older than 1995 which means at least v10.5. A brief version history is available on Wikipedia at [10].

A more precise way to identify the exact tool version is to take the C and C++ runtime libraries from the various compiler revisions, the library files that are linked into the executable like clib3r.lib or plib3s.lib, explode them using wlib tool, disassemble the resulting object files, like undefed.obj, using the wdis tool and compare the signatures of the resulting assembler routines and related variables or lookup tables with the disassembled executable’s routines whether there are exact matches.

workflow for matching library signatures

For the signature checks a set of regex patterns based scripts were developed in Python. The output of the scripts cannot be 100% accurate, manual cross-verification is always required, but this is not a drawback. Having full control over the identification process guarantees that accuracy and completeness are kept high priority. The reversing tool used was IDA Freeware 7.0 which does not have built in signature checkers. The identified routines as well as variables were entered into the IDA database manually, but 100% accurate and as complete as possible with proper type definitions taken from the compiler’s header files.

Detecting the presence of C++ constructs in executables

The Watcom C++32 Runtime library provides services to manage global objects and their life cycles, exceptions and more. The easiest way to detect the presence of C++ constructs in an executable is to search for error messages related to the C++ runtime. A good example is the error message that is emitted when a non existent copy constructor is attempted to be invoked: undefined constructor or destructor called!. There are many more such diagnostic messages.

/*
C++ Runtime Library Error Messages
pure virtual function called!
undefined constructor or destructor called!
compiler error: eliminated virtual function call!
stack data has been corrupted!
violation of function exception specification!
throw while "terminate" function active!
throw during construction of exception!
throw during destructor for handled exception!
re-throw when no exception handler active!
no handler active to catch thrown object!
system exception! code = 0x00000000
no memory left to handle thrown exception!
return from "terminate" function!
return from "unexpected" function!
... 
*/

After basic identification, the previously described signature checks and setup of a reversing database allows identification of the C++ runtime library functions and their calling contexts can be analyzed further.

C++ Runtime Library Function Symbol Name Description
void CPPLIB(fatal_runtime_error)( char *msg, int code ) __wcpp_2_fatal_runtime_error__ Called on fatal runtime errors. Exit the application with diagnostic error message.
void CPPLIB( undefed_cdtor )( void ) __wcpp_2_undefed_cdtor__ The function is emitted by the compiler for undefined constructors and destructors as a placeholder in type signature tables. When a default constructor, copy constructor or destructor is not required by an application they are not generated instead this placeholder function is used.
void CPPLIB(mod_register)( RW_DTREG* rw ) __wcpp_2_mod_register__ Register constructed global, namspace, or class static objects into a linked list to be able to destruct them at termination of the program. These objects are constructed at the start of the program.
void CPPLIB(lcl_register)( RW_DTREG RT_FAR *rw ) __wcpp_2_lcl_register__ Register constructed local static objects into a linked list to be able to destruct them at termination of the program. The main difference compared to CPPLIB(mod_register) is that local static objects are only registered and constructed if their declaration is encountered at least once during program execution.
void CPPLIB(module_dtor)( void ) __wcpp_2_module_dtor__ Iterate the list of previously constructed global, namspace, class static or local static objects in reverse order and call their appropriate destructor.
void * CPPLIB(ctor_array)( void *array, unsigned count, RT_TYPE_SIG sig ) __wcpp_2_ctor_array__ Emitted by the compiler to construct named automatic or temporary object arrays.
void * CPPLIB(dtor_array)( void *array, unsigned count, RT_TYPE_SIG sig ) __wcpp_2_dtor_array__ Destruct a previously constructed object array. The function is not directly emitted into user defined code. For named automic and temporary objects a unique object instance specific function is emitted by the compiler called __arrdtorblk. This instance specific function defines the parameters to be used by the runtime library function.
void* CPPLIB( ctor_array_storage_g )( void* array, unsigned count, RT_TYPE_SIG sig ) __wcpp_2_ctor_array_storage_g__ Emitted by the compiler to register and contruct free-store objects created with the new[] operator.
ARRAY_STORAGE* CPPLIB(dtor_array_store)( void *array, RT_TYPE_SIG sig ) __wcpp_2_dtor_array_store__ Emitted by the compiler before a free-store object array is destroyed with the delete[] operator.
void CPPLIB(pure_error)( void ) __wcpp_2_pure_error__ Trap function for non-overridden pure virtual method calls emitted by the compiler into class virtual function tables.
void CPPLIB(undef_vfun)( void ) __wcpp_2_undef_vfun__ Trap function for stripped virtual function calls. Unless the compiler is broken this function is never called.
   

Finding class members via the this pointer

Named automatic and temporary objects are allocated into automatic memory. In practice this means the stack frame of the function in which the object declaration is found. As constructors, destructors and most class member functions expect the this pointer as the first argument and Watcom typically uses their own register calling convention in these scenarios this creates a promising method to identify class member functions simply by just looking at the assembly listing:

; The %ebp register holds the start address of the stack frame.
; The -0x14 offset from %ebp points to the start address of the object.
; The lea instruction takes the address to the location and moves the value to the %eax register.
; The %eax register holds the first function argument, in this case the 'this pointer'.

lea    -0x000014(%ebp),%eax
call   class_default_ctor_
...
lea    -0x000014(%ebp),%eax
call   class_method_ ; this is clearly a method of the previously identified class.

Finding class default constructors, copy constructors and destructors

Typically the compiler emits class and class instance related data for the C++ runtime into the CONST2 segment. Original C++ module boundaries could be guessed by looking at the layout of this data. For each compiled module the compiler first emits const variable initializers and than the metadata for classes.


By default, the data group “DGROUP” consists of the “CONST”, “CONST2”, “_DATA”, and “_BSS” segments.
The compiler places certain types of data in each segment.
The “CONST” segment contains constant literals that appear in your source code.
The “CONST2” segment contains initialized read-only data.
The “_BSS” segment contains uninitialized data such as scalars, structures, or arrays. [11]


Identification of the default constructor, copy constructor and destructor of global, namspace, class static and local static objects is rather easy in most scenarios.

The compiler emits a type signature for each unstripped class and the data is not stripped even in cases when the C++ runtime does not reference them.

; Comdat: char unsigned const near __typesig[]  SEGMENT ANY 'DGROUP:CONST2'  00000011 bytes  
 0000  00                                      - .
 0001  00 00 00 00                             DD      near A::A()
 0005  00 00 00 00                             DD      near A::A( A const near & ) ; replaced by __wcpp_2_undefed_cdtor__ if function is not referenced
 0009  00 00 00 00                             DD      near A::~A()
 000d  11 00 00 00                             - ....

The first byte is a header, which seems to be always zero in the analyzed executables. In later versions of the compiler the header contains four bytes. The next three 32 bit words are function pointers to the default constructor, copy constructor and destructor in this order. These functions are not necessarily defined for a class. If the class does not have a copy constructor or if it is not referenced, __wcpp_2_undefed_cdtor__ is emitted instead. If the class does not define a default constructor it is replaced with a NULL pointer, but in certain scenarios __wcpp_2_undefed_cdtor__ is emitted instead of NULL. The type signature is not emitted into the executable if the class does not have a default constructor, a copy constructor and the destructor. The last 32 bit word is the object data size in bytes. The given class in the example above has 17 bytes of data. It is very important that alignment rules for basic types do not apply here. In other words object data are always packed.

Global, namspace, and class static objects

For objects constructed, or classes instantiated, at program startup the compiler emits a special function unique for each instance called .fn_init(). This function is responsible to call the appropriate constructors required for the object and track the state of the instantiation.

Debug Options in M.A.X. v1.04

References

[1] Watcom C/C++ compiler options
[2] x86 calling conventions
[3] x87
[4] Timothy Cain biography at MobyGames
[5] HMI SOS home page
[6] Brian Luzietti biography at MobyGames
[7] MVE format
[8] A History of C++: 1979 - 1991, Bjarne Stroustrup - March, 1993
[9] Open Watcom FTP
[10] Watcom C/C++ release history
[11] Open Watcom C/C++ User’s Guide