Kevin Quitt

Program-Transformation.Org: The Program Transformation Wiki
Kevin D. Quitt wrote the following, about his compiler specific approach:

  1. The executable has to be examined to see if it's a compressed executable (like what PKLITE does) and if so, it has to be uncompressed. Or just exclude that type of executable, with information for the user. Check the copyright notice to verify the compiler and version are correct. A single- version decompiler could be expanded to multiple versions of the same compiler by using files containing the critical data for each--but let's work on that later.
  2. Examine the executable for debugging information and retain this information, possibly in a temp file. This can help us detect libraries and recover variable and function names. Build a linked-list that keeps track of what regions of the program are known (fully or partially), and, therefore, which are not known. Each node should have the type of memory (code/data), name of the function/storage if available and type of code (library/user).
  3. Determine the starting point of the executable (past the C startup code), this gives us int main( int argc, char *argv[]" --we're on our way! More importantly, the startup code tells us where initialized variables are and the initial values for them. We don't know, at first, the size or types of these variables; examining the initializing data can help us find floats, doubles, character strings, and maybe int types. By looking for repeating patterns of type/size entries, we can (tentatively) identify arrays of standard types and of structures. We will differentiate between arrays and multiple scalar variables in a row by examining the code that addresses them.
  4. The next step is disassembly: follow the flow of main, noting the addresses of the functions called and the starting addresses of executed instructions. Areas of confusion (where a goto is aimed at the middle of an instruction) must be in-line assembly (and get marked as such) since we always follow program execution. A warning issued when this is seen is probably a good idea; we may be lost already. See if the functions called correspond to anything in the information list. Then, for repeat recursively for each routine called, checking the information list, and comparing the code to the library for that compiler. Any known library modules can be eliminated from further consideration, except when one library function calls another; then we can eliminate the called function(s) as well, etc. By knowing what the various INTs do, we can determine areas that are data. When this process is done (and it may take several passes), we have developed a tremendous information tree about the executable, and we can produce an intermediate assembly-language file that has the library stripped out (even if there was no debugging information), initialized static variables identified along with their initial values.
  5. Once we have the disassembled version, along with our information list, we can start the actual decompilation process. I would use a branching linked list; each node would indicate the name of the function (determined from debug information or synthesized), and a pointer to the code list for that function. The code list node contains the type of code contained (source, binary, or in-line assembly), a pointer to memory that contains the ASCII source, and a pointer to the un-disassembled binary. As code is disassembled, the code list entries split and merge; when the entire function is disassembled, there is a single code list entry. The binary is retained until the entire function has been disassembled.
  6. Concentrate on one function at a time, not following function calls. The entry code to the function tells us about the automatic variables, and examination of the code will reveal the size and probably the type as well. The same type of analysis as for the initialized variable area can be used to locate structs.
  7. The second pass for a function is to identify function calls and the code that passes parameters to the function. For those values passed that are not simple load/push operations, we create a new code node containing the code that generates the value that does get pushed. For functions whose parameter types are know, we can adjust our information about the variables involved (e.g., identifying a "FILE *" variable). This is a form of pattern matching, but should be done before that step to make it easier later.
  8. Pass three is template matching, for whatever templates have been determined for the compiler. As code segments are identified as to the source, a code node is created that adds the generating source code to the binary. Template matching must be done recursively to handle loops within loops, etc. The setup code for loops either identifies them as for, while, or do, or the loop type doesn't matter and we can choose arbitrarily. This will pick up switches, cases, ifs, and function calls turned into in-line code by "#pragma inline" statements. When #pragma controlled code is encountered, the appropriate #pragma statement must be put in the source; the actual location depends on whether the #pragma can be used is a local block of code, or only for the entire file (and is there a corresponding #pragma that turns it off).
  9. Pass four looks for in-line code that has a direct source representation and adding the source to the code node. This is code that picks up one or more variables, does something to them, and puts one or more values back. Attention must be paid to the sources of register variables (from the automatic pool) so that source can correctly identify the original variable. The conditional in "if" statements can be determined at this point since we can now identify the (immediate) source of the comparison.
  10. Anything left over at this point can probably be safely turned into in-line assembly code. In any case, that's the ultimate fall-back position.

Although he thinks that dcc will be better than exec-2-c, he doubts whether their approach will lead to better results than his, both with respect to reconstructing the original work-flow, and with respect to reconstructing the data type information. He commented: In theory, theory is the same as practice--in practice, it's not.

Kevin worked on this in the past for his own, and is perfectly willing to discuss the decompilation process with anybody and tell them his ideas and suggestions for various techniques and processes. He has some scattered programs that he used to test out individual ideas, but does not have a set of files that he would call `source for a decompiler'.