Decompiler For Pseudo Code Generation

San Jose State University San Jose State University

SJSU ScholarWorks SJSU ScholarWorks

Master's Projects Master's Theses and Graduate Research

Spring 2011

Decompiler For Pseudo Code Generation Decompiler For Pseudo Code Generation

Ankit Patel

San Jose State University

Follow this and additional works at: https://scholarworks.sjsu.edu/etd_projects

Part of the Other Computer Sciences Commons, Programming Languages and Compilers Commons,

and the Theory and Algorithms Commons

Recommended Citation Recommended Citation

Patel, Ankit, "Decompiler For Pseudo Code Generation" (2011).

Master's Projects

. 182.

DOI: https://doi.org/10.31979/etd.j55q-fr8s

https://scholarworks.sjsu.edu/etd_projects/182

This Master's Project is brought to you for free and open access by the Master's Theses and Graduate Research at

SJSU ScholarWorks. It has been accepted for inclusion in Master's Projects by an authorized administrator of SJSU

ScholarWorks. For more information, please contact [email protected].

i

Decompiler For Pseudo Code Generation

A Project Report

Presented to

The Faculty of the Department of Computer Science

San Jose State University

In Partial Fulfillment

Of the Requirements for the Degree

Master of Computer Science

By

Ankit Patel

May 2011

ii

Ankit Patel

SAN JOSÉ STATE UNIVERSITY

iii

The Undersigned Project Committee Approves the Project Titled

Decompiler For Pseudo Code Generation

By

Ankit Patel

APPROVED FOR THE DEPARTMENT OF COMPUTER SCIENCE

___________________________________________________________

Dr. Mark Stamp, Department of Computer Science Date

___________________________________________________________

Dr. Robert Chun, Department of Computer Science Date

___________________________________________________________

Ravi Savaliya, Senior Software Developer, Ebay Date

______________________________________________________________

Associate Dean Office of Graduate Studies and Research Date

iv

Abstract

Decompiler For Pseudo Code Generation

By Ankit Patel

Decompiling is an area of interest for researchers in the field of software reverse

engineering. When the source code from a high-level programming language is compiled, it

loses a great deal of information, including code structure, syntax, and punctuation.

The purpose of this research is to develop an algorithm that can efficiently decompile

assembly language into pseudo C code. There are tools available that claim to extract high-level

code from an executable file, but the results of these tools tend to be inaccurate and unreadable.

Our proposed algorithm can decompile assembly code to recover many basic high-level

programming structures, including if/else, loops, switches, and math instructions. The approach

adopted here is different from that of existing tools. Our algorithm performs three passes through

the assembly code, and includes a virtual execution of each assembly instruction. We also

construct a dependency graph and incidence list to aid in the decompilation.

v

ACKNOWLEDGEMENTS

I would like to thank Dr. Mark Stamp, for guiding me through this research project and

working with me to achieve this. I also thank him for his suggestions and contributions for

handling some of the difficulties faced during the course of this project. Without him, this would

not have been possible.

I would also like to thank Prof. Debra Caires for teaching me an efficient and effective

thesis documentation process.

vi

Table of Contents

1. Introduction ............................................................................................................................................... 1

1.1 Objective ............................................................................................................................................. 2

2. Background ............................................................................................................................................... 3

2.1 Fundamentals of forward and reverse engineering ............................................................................. 3

2.2 Windows portable executable file format ........................................................................................... 5

2.3 Assembly language ............................................................................................................................. 6

3. Decompilation ........................................................................................................................................... 9

3.1 Benefits of Decompiling ..................................................................................................................... 9

3.2 Decompiler problems ........................................................................................................................ 11

3.2.1 Information loss ......................................................................................................................... 11

3.2.2 Separation of .code and .data sections....................................................................................... 13

3.2.3 Differentiating original pointers from address offsets ............................................................... 14

3.2.4 Detecting constants from the pointers........................................................................................ 14

3.3 Limitations ........................................................................................................................................ 15

3.4 Types of Decompilers ....................................................................................................................... 15

3.5 Present decompiler tools ................................................................................................................... 17

5. Decompiler algorithm ............................................................................................................................. 20

5.1 Initial Preparation .............................................................................................................................. 20

5.2 Algorithm .......................................................................................................................................... 22

5.1 First Pass – Raw view ....................................................................................................................... 24

5.1.1 Raw view example ...................................................................................................................... 24

5.2 Second Pass – Parsed view ............................................................................................................... 25

vii

5.2.1 Parsed view example .................................................................................................................. 27

5.3 Third Pass – Recognized view .......................................................................................................... 27

5.3.1 Stack initialization and analysis ................................................................................................ 28

5.3.2 Register initialization and analysis ............................................................................................ 29

5.3.3 Graph Generation ...................................................................................................................... 30

5.3.4 Output generation ...................................................................................................................... 38

5.3.5 Output ........................................................................................................................................ 39

6. Test cases and Results ............................................................................................................................. 40

6.1 Comparison with manual decompilation .......................................................................................... 42

8. Future Work ............................................................................................................................................ 45

Appendix A – Test cases and results .......................................................................................................... 47

References ................................................................................................................................................... 59

List of Figures

Figure 1: Forward and reverse engineering .................................................................................................. 4

Figure 2: Assembly instruction format .......................................................................................................... 8

Figure 3: Assembly code snippet of an algorithm checking the number of free trails used in the software

.................................................................................................................................................................... 10

Figure 4: Decompiling algorithm flow chart ............................................................................................... 23

Figure 5: Raw view input ............................................................................................................................. 24

Figure 6: Parsed view output ...................................................................................................................... 27

Figure 7: Stack Initialization ........................................................................................................................ 28

Figure 8: Registers initialization .................................................................................................................. 29

Figure 9: Nested if dependency graph ........................................................................................................ 32

Figure 10: If - then dependency graph ........................................................................................................ 33

viii

Figure 11: Dependency graph for if/else .................................................................................................... 34

Figure 12: Dependency graph for loop ....................................................................................................... 35

Figure 13: Dependency graph for switch case ............................................................................................ 37

Figure 14: Recognized view output ............................................................................................................. 39

Figure 15: Decompiler generated output ................................................................................................... 42

Figure 16: Decompiler generated output ................................................................................................... 44

List of Tables

Table 1 : High level code and assembly code comparison............................................................................ 1

Table 2: Examples of high level code and assembly code ............................................................................ 2

Table 3: Comparison of Dcc and REC .......................................................................................................... 19

Table 4: Assembly instruction string format ............................................................................................... 24

Table 5: Supported assembly instructions .................................................................................................. 26

List of Code

Code 1: C code for understanding read/write access of .data section ......................................................... 5

Code 2: Unoptimized high level code ......................................................................................................... 11

Code 3: Code before compilation ............................................................................................................... 12

Code 4: C code example containing data within code ................................................................................ 13

Code 5: Code example for pointer offset outside the bounds of array ...................................................... 14

Code 6: Code example for pointer offset inside the bounds of array ........................................................ 14

Code 7: Function calling convention ........................................................................................................... 21

Code 8: Recognized instruction C structure ................................................................................................ 25

Code 9: Supported instructions' function handlers .................................................................................... 31

Code 10: Nested if ....................................................................................................................................... 31

ix

Code 11: Test case C code ........................................................................................................................... 40

Code 12: if/else original C code .................................................................................................................. 43

Code 13: Manually generated output ......................................................................................................... 43

List of Assembly code

Assembly code 1: Example of assembly code depicting information loss .................................................. 12

Assembly code 2: Loss of data in .Code section ......................................................................................... 13

Assembly code 3: _stdcall calling convention ............................................................................................. 21

Assembly code 4: _cdecl calling convention ............................................................................................... 22

Assembly code 5: _fastcall calling convention ............................................................................................ 22

Assembly code 6: Parsed view input........................................................................................................... 27

Assembly code 7: Test case assembly code ................................................................................................ 41

Assembly code 8: Assembly code if/else .................................................................................................... 43

1

1. Introduction

“Reverse engineering is the process of identifying a system's components and their

interrelationships, and creating representations of the system in another form or at a higher

level of abstraction.” [9]

Source code is compiled to assembly code by the compiler. Assembly code consists of a series of

instructions that is executed by the micro-processing unit of the computer [7], whereas source

code is usually written in high-level programming languages like C, C++, Java, or C# [23].

These languages are designed to be used and understood by humans to program a computer.

Software requires constant maintenance and upgrades to deliver the best performance, user

experience, and functionality, and high-level languages play an important role in understanding

and modifying software. A decompiler comes into play when the source code is not available

[14] [4].

Complete decompilation of assembly code is not only difficult, but very limited [1] [3]. High-

level programming language is very detailed and descriptive, following specific structures and

syntax that make it very easy to understand – the programmer does not need to worry about how

the hardware will execute the code [32]. On the other hand, assembly language is significantly

hardware-dependent; assembly code is series of instructions to be performed by the processor,

and hence it does not need structure and syntax [8]. Table 1 shows some of the differences

between high-level language and assembly code [1].

Table 1 : High-level code and assembly code comparison

High-level code Assembly code

Highly structured Less structured (series of instructions)

Complex expressions Basic expressions only

Machine independent Highly machine-dependent

Low detail High detail

High level Low level

Table 2 shows examples of high-level code and assembly code, illustrating Table 1:

2

Table 2: Examples of high-level code and assembly code

C code Assembly code

void test ( int a, int b )

{

if ( a < b )

printf ( "a < b\n" );

}

011411A0 push ebp

011411A1 mov ebp,esp

011411A3 sub esp,40h

011411A6 push ebx

011411A7 push esi

011411A8 push edi

011411A9 mov eax,dword ptr [ebp+8]

011411AC cmp eax,dword ptr [ebp+0Ch]

011411AF jge 011411BF

011411B1 push 1145720h

011411B6 call dword ptr ds:[1148248h]

011411BC add esp,4

011411BF pop edi

011411C0 pop esi

011411C1 pop ebx

011411C2 mov esp,ebp

011411C4 pop ebp

011411C5 ret

1.1 Objective

This research paper deals with the decompilation of assembly code to pseudo C code.

Decompiling plays an important role in the field of software reverse engineering because often a

software developer needs to understand the assembly language produced from the source code.

For example, there are critical security features of software that are vulnerable at assembly level,

and hackers can easily exploit them [1]. Studying the code at the assembly level can help the

developer to implement security features more efficiently, and can also help in software

reusability [9]. But, as mentioned earlier, assembly language is not casually readable – it is

extremely time-consuming to understand even a small piece of code.

The algorithm proposed in this paper generates a pseudo C code from an input assembly code.

The initial preparations for this algorithm included generation of sample assembly code from

small C codes using the Microsoft Visual C++ compiler. These test cases were designed to cover

most of the high-level programming language syntaxes and structures, such as if/else, switch

case, and loops. The assembly code generated from these sample source codes was analyzed for

the critical and commonly-generated assembly instructions. These assembly codes were then

stored in a .dis extension file, which is actually a text file. The algorithm proposed in this paper

consists of three modules, which are discussed in later sections. The assembly code passes

through each of these modules to produce a pseudo C code in the output.

3

2. Background

To understand the purpose of this paper, it is crucial to understand the fundamentals of forward

and reverse engineering, Windows portable executable (PE) file format, and assembly language.

The discussion of forward and reverse engineering explains the different phases of software

during the compilation and decompilation process. The Windows portable executable file format

section explains the structure and sections of the PE file format. And finally, the assembly

language section explains the x86 assembler, assembly instruction format, and different

categories of assembly instructions.

2.1 Fundamentals of forward and reverse engineering

In forward engineering, source code passes through four phases: compiling, assembling, linking,

and execution [32]. Reverse engineering deals with these four phases in reverse order –

execution, linking, disassembling, and then decompiling. A considerable amount of information

is lost in the transition through these phases, and is unrecoverable in the reverse transition [1].

Reverse engineering consists of many practices, such as reverse assembling from native machine

code (disassembling), reverse compiling from assembly code (decompiling), reverse

programming from the source code itself (debugging), reverse programming of legacy code, and

software reusability [3]. This paper deals with the implementation of reverse compiling from

assembly code, a process that faces the most difficulties compared to other techniques. Most

compilers generate the assembly code from the source code, which is then parsed by the

assembler to generate the object code [32]. These two phases are replaced by the interpreter in

scripting languages like Perl, PHP, and JavaScript, as they generate the object code directly from

the source code [23]. Virtual compilers (for example, Java compiler) produce byte code, which is

equivalent to the object code.

Figure 1 shows the relationships between the various tools used with forward and reverse

engineering [1]. The extreme right indicates the tools that take the software from one phase to

another. The nodes indicate the language or level of the code at that particular phase. The arrows

from top to bottom relate to the software life-cycle in forward engineering, and from bottom to

top in reverse engineering. This paper deals with the phase indicated by the arrow in purple.

4

Figure 1: Forward and reverse engineering

Program

Source

codes

Assembly

language

Object

code

Machine

code

End User

Application

Consists

IDE integration

Compile

Assemble

Link

Execute

Disassemble

Decompile

Advanced Decompile

Object code decompile

COMPILER

ASSEMBLER

LINKER

OPERATING SYSTEM

Forward

Engineering

Achievable

LEGENDS

FLOW

Difficult & Limited

Reverse

Engineering

Execute

Compile

IDE

5

2.2 Windows portable executable file format

Portable executable (PE) is a Microsoft-defined file format for the executables in the Windows

operating systems. It contains the required information for a loader to manage executable code in

the windows environment [12]. The linker gets its information regarding sections and headers

from the PE file – the information contained in the PE file decides how the linker loads the file

into memory during execution. PE consists of two types of section, the .data section and the

.code section (also known as .text section) [12]. Each section possesses its own memory

attributes that define whether data is shared between different processes, whether the section

contains code, and read–write access. It also consists of more information stored in headers, such

as the PE header, DOS header, and section table, which are required for the execution of the

program [12]. However, that information is not relevant to our research, and so is not discussed

here.

.data section

The .data section is one of the very important sections of the PE file format, containing

information about the static and global variables that were initialized in the source code. It is a

fixed-size section – like the .code section in the PE file – because the information stored in the

.data section is defined by the programmer before the code is compiled. All of the constants and

static variables used in the source code are stored in the .data section [12]. The .data section also

contains resources and relocations, API imports, and exports tables. The .data section is not read

because the values stored in the variables are likely to change during the execution. The example

below explains this phenomenon:

int i[] = {1,2,3,4,5};

int j = 6;

void main(){

char * test = “Hello World”;

int x = 5;

int* k = &x;

}

Code 1: C code for understanding read/write access of .data section

6

Integers i and j defined outside of the main will be stored in the data section as read-write,

whereas the string literal, “Hello World,” and integer value 5 will be stored as read only. The

pointer variables test and k will be stored in the read-write area.

.code section (.text section)

The .code section is a crucial section of the program that contains executable assembly

instructions. It is a fixed-size section, but – unlike the .data section – the .code section is read

only. However, the architecture supporting the self-modifying code can remove the read-only

constraints of this section, and this functionality allows for virus programming [1]. The position-

independent code of this section can be shared in memory across various processes. This section

is the most interesting section in reverse engineering, as it contains the actual logic and code

behind the software [14]. Hence, it becomes the source of input for the decompilation process.

The input to the algorithm discussed in this paper is the code section extracted from the PE.

2.3 Assembly language

The decompiling process starts from the machine code or the assembly code, with source code as

the final output. Even if it is a machine code decompiler, decompiling from machine code to

assembly code is much simpler than the second stage, which is the conversion of assembly code

to high-level language [4]. Hence, this section deals with the basic introduction of the assembly

language and x-86 instruction set.

Assembly language is a low-level language for programming hardware like the CPU, integrated

chips, and microcontrollers [7]. An assembly language is simply a machine-specific and non-

portable symbolic representation of the underlying machine code that can be interpreted by a

machine [7]. The symbols are called mnemonics, and are designed by the manufacturer of the

hardware. An assembler is basically a utility that does the same job as a compiler does for high-

level languages [8] – it translates the mnemonics into respective binary sequences. Higher-level

assemblers, like those installed in computers, convert assembly instructions into object code

rather than machine code. The object code consists of opcodes for all assembly instructions,

instead of the binary sequence [8]. This approach is adopted because of the large amount of

7

instruction sets for computers compared to other devices. Some of the notable assemblers are

MIPS, SPARC, and x86. This paper deals with the latter, x86.

x86 Assembler

This assembler is based on the 8086 assemble architecture of the Intel processor. It is supported

by Intel, AMD, and VIA processor-based systems. Operating systems like Windows, Linux, and

MAC-OSX support this assembler as well. Two versions – x86(32) 32-bit, and x86(64) 64-bit –

are available based on the processor cores [8].

x86 performs two passes on an assembly code. The first pass parses the source code to create a

table with all the unresolved symbols. The second pass uses this information to resolve the

address. This architecture lets you use undeclared symbols in your code, as opposed to one-pass

assemblers which do not [8].

x86 Instruction Set

The x86 instruction set is designed for Intel and AMD processors, and is backwards-compatible

with the following previous versions designed by Intel: 80186, 80286, 80386 and 80486 [8].

x86 processors use different registers for storing values [7]. They are,

• Special registers AX, BX, CX, and DX

• IP (Instruction Pointer)

• Flags like carry, sign, zero, and many others.

• Segment registers CS, DS, ES, FS, GS, and SS, for representing the six different

sections, such as Code, Data, Stack, and Extra.

x86 instructions are divided into five different categories based on their functionality:

1) Stack Instructions

• These instructions deal with the stack manipulation of the processor.

• Instructions like PUSH, POP, CALL, and RET come under this category.

2) Integer ALU Instructions

• These instructions deal with mathematical calculations, logical operations, and

shift operations.

8

• Arithmetic instructions: ADD, SUB, and MUL.

• Logical instructions, such as XOR, AND, and OR.

• Shift instructions: ROL, ROR, RCL, and RCR.

3) Floating Point Instructions

• These instructions deal with more complex mathematical operations, like

square root, division, and modulo, with floating point.

4) SIMD Instructions

• These instructions can perform several calculations in parallel in SIMD

registers. These are supported by modern x86 processors only.

5) Data Manipulation Instructions

• These instructions are used for accessing the data in the memory from the

different sections using various addressing modes.

• MOV, ENTER, and LEAVE

Assembly instruction format

Each line of code in the assembly language is one instruction. This pass reads it as a string. All

assembly instructions follow a specific format, unlike high-level programming language. This

format consists of four sections that are common throughout the whole assembly code:

Figure 5.1: Assembly Instruction Format

Each string contains,

1) Memory location

2) Assembly instruction

3) Operand 1 (optional)

4) Operand 2 (optional)

Figure 2: Assembly instruction format

9

3. Decompilation

In the previous chapter we discussed the fundamentals of forward and reverse engineering.

Decompilation is a process of reverse engineering, and the main focus of this research. This

chapter focuses on the benefits and problems of decompiling, and the types of decompilers and

present available tools. Decompilation is, and always will be, unable to achieve a 100% success

rate, because of many factors discussed in later sections [1] [3] [4]. There are many different

compilers available today for a programming language [32][23]. Each is implemented to handle

the source code differently in order to produce the most optimized assembly code for the

underlying machine. In addition, high-level programming languages are growing faster and

faster compared to their origins, with the inclusion of more libraries to build complex software to

satisfy new and growing requirements [23].

The most important requirement expected from a decompilation process is the human readability

of the decompiled output [3]. Often, the assembly code is too difficult to understand even for the

developer of the source code himself. Even the assembly code of a small application like “Hello

World” would take a few minutes to understand. The output of this process should be much

smaller than the assembly code, and should not contain any irrelevant details of the hardware,

memory access, or operating system interactions. There is not a tool, yet, to achieve all of these

requirements [1].

3.1 Benefits of Decompiling

Even if the results of the decompiling process are not fully accurate to the original source code,

there are some benefits to this process as well. These benefits apply to three categories,

depending upon their application [1].

1) Analyzing

2) Error checking and Evaluation

3)

Updating and Optimization

10

Analyzing

Reverse engineered source code can be used to learn the underlying algorithm and design

principles incorporated in the program [1]. This information can be used to develop a new code

with a better design and more efficient algorithm. Even if the software uses a very powerful

security algorithm at high-level programming, the most complex part of the algorithm when

compiled to assembly code is converted to a series of granular-level instructions [9]. To

overcome this security, a hacker needs to change only a very few instructions. As in this

example, by only changing one assembly instruction the hacker could make the number of free

trials offered by the software unlimited.

Figure 3: Assembly code snippet of an algorithm checking the number of free trials used in the software

The decompiler can help the developer to understand this assembly language, and therefore to

program accordingly at high level. The developer can make the required changes to protect the

vital information in the assembly code.

Error checking and Evaluation

There are several bugs that remain undetected in high-level programming languages that can be

discovered by a careful analysis of the registers and stacks while debugging the assembly

language [10]. Infusing a viral code into the empty sections of an executable is a very common

method of attack used by virus developers around the globe. The decompilation process can help

distinguish foreign code from the original source code. A good example of this would be Cheat

Engine, an auto-assembler tool for injecting code into running processes that supports the x86

assembly instruction set. Cheat Engine is used in multiplayer gaming clients to manipulate player

levels and awards.

An ideal decompiler would produce the identical source code of two different assembly

languages having the same logic [4]. This feature could be used to detect copyright

11

infringements. Other applications for the decompiler include bug fixes, finding vulnerabilities,

interoperability, signature detection, and code comparison.

Updating and Optimizing

When source code is compiled, the compiler performs a number of optimizations on the code to

achieve the best performance from the machine. The reversed source code can be studied to

understand the optimizations performed by the compiler [1]. This information is helpful in

optimizing the code in high-level programming language itself. Also, possible modules that need

new updates and optimizing can be checked. Considering the following code:

if (4>5){

printf (“This will never execute”);

}

Code 2: Unoptimized high-level code

The compiler will remove the above code when generating the assembly code, as this if

statement is logically incorrect and execution will never print the statement.

3.2 Decompiler problems

Most of the decompiling tools in use today face common problems. In this section, we discuss

the major problems of this process, considering C as both the original source language and the

target language. These problems make the decompilation process impossible because of the way

high-level languages are structured. Overcoming such problems requires a change in the process

of forward engineering.

3.2.1 Information loss

The compiler performs code optimization, semantic analysis, lexical analysis, code generation,

and preprocessing. As a result, a great deal of information embedded in the source code is lost.

This information can be manually handled, but reversed code loses the originality and feel of the

source code. Information lost during compilation is as follows:

1) Complex mathematical expressions

12

2) Global and local variable names

3) Object-oriented code

4) High-level programming syntax

5) Structural programming

6) Comments

7) Data types

The below example consists of original C code and compiler-generated assembly code. It

demonstrates the substantial loss of information in compilation described above.

Original C code

void __cdecl Test2a ( int a, int b )

{

if ( a < b )

printf ( "a < b\n" );

else

printf ( "a >= b\n" );

}

Code 3: Code before compilation

Assembly code

01141250 push ebp

01141251 mov ebp,esp

01141253 sub esp,40h

01141256 push ebx

01141257 push esi

01141258 push edi

01141259 mov eax,dword ptr [ebp+8]

0114125C cmp eax,dword ptr [ebp+0Ch]

0114125F jge 01141271

01141261 push 1145720h

01141266 call dword ptr ds:[1148248h]

0114126C add esp,4

0114126F jmp 0114127F

01141271 push 1145728h

01141276 call dword ptr ds:[1148248h]

0114127C add esp,4

0114127F pop edi

01141280 pop esi

01141281 pop ebx

01141282 mov esp,ebp

01141284 pop ebp

01141285 ret

Assembly code 1: Example of assembly code depicting information loss

13

3.2.2 Separation of .code and .data sections

A high-level programming language consists of two sections, .code and .data. Both are

seamlessly merged throughout the code. Consider a code section as the core logic involving math

operations, and a data section as variable values, strings, and constants [10]. When compiled, the

assembly code separates this information into two different sections before embedding it into a

portable executable. The data section is then referred to by the code section of the assembly

through the memory address where this data is stored. It is difficult to merge these two sections

to their original form during the decompilation process [1] – the unavailability of the

corresponding data of the code section requires assumptions in the process, and hence the output

generated is not reliable. These problems arise because C follows an approach of defining all

items before using them. The compiler then knows what to extract from the source code for the

data section.

Consider the following C code, in which the data contained in the string passed to printf is

unavailable in the code section of the respective assembly language. This function call is

converted to call dword ptr ds:[1148248h]

in the assembly code.

Original C code

void test ( int a, int b )

{

if ( a < b )

printf ( "a < b\n" );

}

Code 4: C code example containing data within code

Assembly code

01141210 push ebp

01141211 mov ebp,esp

01141213 sub esp,40h

01141216 push ebx

01141217 push esi

01141218 push edi

01141219 mov eax,dword ptr [ebp+8]

0114121C cmp eax,dword ptr [ebp+0Ch]

0114121F jge 0114122F

01141226 call dword ptr ds:[1148248h]

0114122C add esp,4

0114122F pop edi

01141230 pop esi

01141231 pop ebx

01141232 mov esp,ebp

01141234 pop ebp

01141235 ret 8

Assembly code 2: Loss of data in .code section

14

3.2.3 Differentiating original pointers from address offsets

Assembly language is ambiguous in differentiating between the actual pointer variables – used to

store an address or the starting address of a dynamic memory allocation in a high-level code –

and the offset pointers that point to an address offset from the actual pointer [1].

Consider the following 2 cases:

Pointer offset outside the bounds of array

int numbers [5] = {1, 1, 1, 1, 1};

for(int i = -5, i < 0; i ++){

printf(“%d”, a + i);

}

Code 5: Code example for pointer offset outside the bounds of array

Pointer offset inside the bounds of array

int numbers [5] = {1, 1, 1, 1, 1};

for(int i = -5, i < 0; i ++){

printf(“%d”, a + i + 5);

}

Code 6: Code example for pointer offset inside the bounds of array

This type of code produces the same assembly output, but when reverse engineered only the

second case will produce the closest corresponding C code to the original source code [2]. This is

because the second case is still logically inside the array index bounds, while the first case is not

at all accessing the array elements [23]. Without the compiler and linker differentiating between

the original pointer and the offset, it is impossible to determine the offset for a reverse

engineering tool just from the pointer variable [32].

3.2.4 Detecting constants from the pointers

The other major problem faced by reverse engineering tools is distinguishing pointers from

constants [1]. The operand following the assembly instruction is treated as a constant, which can

be any data type – an integer or a float. But, it can also be a pointer variable to a data stored

15

somewhere in the memory instead of the data itself. This problem can be considered as the first

hurdle towards differentiating offsets and actual pointers. Using actual data or an address from a

.data section that contains the data is decided by the compiler based on the optimization setting

[8].

3.3 Limitations

Though decompilers can be used for a variety of benefits, achieving those benefits requires

considerable manual intervention and numerous assumptions during the process. These

assumptions can sometimes lead to a faulty code and an even more complex structure. Also, if

time is a constraint, manual intervention is impossible for larger codes. The primary existing

limitations that need to be addressed while designing a decompiler are as follows: [1]

• Handling of direct or indirect function calls

• Handling of direct or indirect jump calls

• Handling type casting and recovery

• Stack pointers or memory references

• Function and variable names

• Merging code and data sections

The only advantage of reverse engineering over forward engineering is its full scope over the

entire application.

3.4 Types of Decompilers

Decompilers can be summarized into three different types. Each has its own difficulty level and

constraints. These difficulties arise based on the desired input and output [4], while constraints

are present due to the information loss and the optimizations performed by the compiler. Tools

that reverse the actions of linkers and disassemblers are currently available, with highly-

dependable results. IdaPro and OllyDbg are considered to be the best of them [4].

Many attempts and much research has suggested the use of different decompilers. However,

these tools are unable to produce results that can be compiled or even be understood by humans

[3][4]. This section categorizes different decompiler tools and gives a few examples of each. The

16

most important of these, and the one that this paper deals with, are machine code decompilers. A

more detailed explanation with results analyzed from the currently-available machine code

decompilers is also presented.

Virtual machine decompiler

A fine example of the virtual machine decompiler is the Java bytecode decompiler. The Java

bytecode decompiler has achieved a success ratio near 90%, as the Java byte is rich with

metadata to allow negligible information loss. Most of the complex information – such as

variable names, separation of code and data section, type analysis, and global data – is explicitly

visible in bytecode [1].

There exist a few challenges with the virtual machine decompiler, too, but these have already

been resolved to an extent. Optimized bytecode, code obfuscation, and subtypes of basic

datatypes are a few of these difficulties.

Virtual machine decompilers enjoy the best position because their input is rich with metadata,

which is vital for the decompiling process. Here is a list of some professional virtual machine

decompilers:

1) JAD Decompiler [28]

2) JReverse Pro [29]

3) JODE Decompiler [27]

4) McGill’s Decompiler [26]

Object code decompiler

An object code decompiler takes the machine code and generates the intermediate object files

between the machine code and assembly code [31]. Though theoretical only, these decompilers

are better than machine code decompilers but worse than assembly code decompilers in terms of

achievability [1] [3]. The object code decompiler does not benefit from much research focus, as

its final output is object code that serves no major importance in reverse engineering [1]. As

such, very few of these decompilers have been researched. The most noteworthy object code

decompilers are Winger and Schneider [30] and Decomp [31]. Both were created with very

limited functionality and for specific purposes only.

17

The machine code decompiler discussed in the next section is a multi-process decompiler. One

of its processes has the same functionality as an object code decompiler – reversing the machine

code – but the final output of a machine code decompiler is assembly code, not object code [4].

This process is beyond the scope of this paper and not discussed in detail.

Machine code decompiler

Machine code decompilers have attracted the most research from enthusiast reverse engineers

since the origin of programming language. Since these decompilers are based directly on the

underlying platform (that is, hardware), these are the most difficult decompilers to create. A

universal decompiler is also not achievable even in theory [1]. Even with the constant evolution

of high-level programming languages and the hardware that processes machine code, these

decompilers still have not reached a level to be able to extract even 10% of the original source

code [32]. The two processes of the machine code decompiler are disassembling and

decompiling, but we concentrate on the decompiling process here. A few noteworthy research

machine code decompilers are listed below. The first three are the most successful, and are

discussed in more detail later in this paper.

1) Boomerang [1]

2) Dcc [5]

3) REC [6]

4) Exec -2-C [20]

5) DisC [25]

6) DesQuirr [16]

7) Yadec [17]

8) Andromeda [18]

9) HexRays [19]

3.5 Present decompiler tools

For decompiling a portable executable compiled with most commonly-known compilers –

Borland C++, Dev C++, Microsoft Visual C++, and Turbo C – there are very few decompilers

available that perform this process with considerable success ratio [3][4]. These are Boomerang,

REC, and dcc [3]. There have been many other decompiler tools prior to this, but they were more

18

like research tools that perform only in a controlled environment. They are classified based on

their category in the next chapter. As we are interested in C code generation, only these three

most functional tools match our criteria.

Boomerang

Boomerang is an open-source machine code to C decompiler. It tries to alter the semantics of

each of the assembly instructions, and implements the Static Single Dataflow Analysis to achieve

its goal [1]. This makes it independent of the compiler optimizations performed on the assembly

code. Boomerang’s approach to this process includes implementing a very powerful internal

representation (IR), which then tries to recreate each and every step of the compiler while

parsing the assembly code. However, the results on the Hello World program were either

inefficient or not even close to the original, and the process is quite time-consuming. The

algorithm described in this paper adopts these concepts from Boomerang, but they are modified

to produce better results. The internal representation of symbols is critical for implementing a

decompiler algorithm, as the assembly code is very granular and we cannot be sure about each

symbol as we encounter them in passes.

REC

REC is a machine code decompiler that recognizes not only Win32 executables but also other

formats of Linux, Solaris, and Playstation PS-X executables. The output generated is C-like, and

requires manual editing to recompile [5]. Though the output produced by REC is very close to C

code, the decompiler has many constraints. The REC decompiler is heavily compiler-dependent,

supporting Win32 executables (portable executables) compiled by Microsoft Visual C++ 6 and

Microsoft Visual Basic 5. The algorithm relies on the information available from the executable

symbol table, which varies per compiler settings. It performs poorly if the executable is created

without debugging information files like the program data base (.pdb) or the code view (.c7).

Dcc

Dcc is a research decompiler that decompiles only 80286 DOS-based executables to C language

[6]. The Dcc approach to decompilation is based on graph theory and optimization techniques

adopted by compilers. These compiler optimization techniques eliminate high-level

19

programming constructs and generated intermediate assembly instructions using registers. Dcc

tries to follow the same approach in a reverse order. Dcc is also a three-pass decompiler, where

the first pass is a machine code decompiler that generates assembly, the second pass generates

intermediate symbol tables by performing code flow and data flow analysis, and the third pass

generates the output, which resembles C like code. The algorithm described in this paper follows

a three-pass, but the functionalities implemented at each pass are completely different from those

of this tool. The machine code decompiler is beyond the scope of this research, but graph

representation of the control flow is adopted in this paper.

Comparison of Dcc and REC

The table below compares different aspects of the two best available decompilers on a scale of 0

to 10, with 0 being the worst performance and 10 being the best [1].

Table 3: Comparison of Dcc and REC

Test Case Dcc REC

Large Executables 5

0

Parameters Handling 7

5

Handling Returns 7

5

Jumps 9

5

Function Calls 7

0

Type casting 7

0

Total 42

15

20

5. Decompiler algorithm

This section deals with the main algorithm behind the decompiling process implemented in this

paper. The algorithm is a three-pass algorithm: raw view, parsed view, and recognized view.

Raw view deals with reading the assembly language from the .DIS file. This output is still very

difficult to handle within the algorithm. Parsed view generates the vector of strings in a specific

format that can be understood by the algorithm. The third pass, recognized view, contains the

algorithm that reads one assembly line at a time and generates the pseudo C code. The algorithm

is implemented in C language using the Microsoft Visual C++ compiler. Code optimization

options of the compiler are disabled.

5.1 Initial Preparation

A wide range of test functions – including different implementations of C basic syntax like for

loop, while loop, if/else (simple and complex), switch case, and math instructions – were created

using Microsoft visual C++ compiler. Compiler optimizations were disabled for creating these

test cases, as it is trivial for the algorithm to have the complete assembly of the logic, which may

contain assembly instructions that would otherwise have been omitted by the compiler

optimization.

Calling Conventions

A calling convention is used to determine how different platforms parse different functions, how

their parameters are received and handled, and how results are returned from the function.

Different programming languages use different calling conventions. All test cases were created

using three different calling conventions: _stdcall, _cdecl, and _fastcall. These calling

conventions are provided by the Microsoft Visual C++ compiler, and are considered in this

research because they play an important role in how the assembly code handles parameters, stack

operations, and registers. The example below explains different calling conventions in detail.

First consider the below C code where XX can be substituted with _stdcall, _cdecl and, _fastcall

keywords.

21

Original C code

void XX Test1b ( int a, int b )

{

if ( a < b )

printf ( "a < b\n" );

}

Code 7: Function calling convention

1) _stdcall

Parameters passed to the function are pushed to the stack in reverse order. The function

accesses these parameters from the stack, and the stack cleans operations after the

function execution is handled by the callee.

01141210 push ebp

01141211 mov ebp,esp

01141213 sub esp,40h

01141216 push ebx

01141217 push esi

01141218 push edi

01141219 mov eax,dword ptr [ebp+8]

0114121C cmp eax,dword ptr [ebp+0Ch]

0114121F jge 0114122F

01141221 push 1145720h

01141226 call dword ptr ds:[1148248h]

0114122C add esp,4

0114122F pop edi

01141230 pop esi

01141231 pop ebx

01141232 mov esp,ebp

01141234 pop ebp

01141235 ret 8

Assembly code 3: _stdcall calling convention

2) _cdecl

This calling convention is the same as _stdcall, except that the stack cleanup operations

are performed by the caller. It is clearly visible in the example below that the function

does not return anything, as the results are stored in the stack.

011411A0 push ebp

011411A1 mov ebp,esp

011411A3 sub esp,40h

011411A6 push ebx

011411A7 push esi

011411A8 push edi

011411A9 mov eax,dword ptr [ebp+8]

011411AC cmp eax,dword ptr [ebp+0Ch]

011411AF jge 011411BF

011411B1 push 1145720h

22

011411B6 call dword ptr ds:[1148248h]

011411BC add esp,4

011411BF pop edi

011411C0 pop esi

011411C1 pop ebx

011411C2 mov esp,ebp

011411C4 pop ebp

011411C5 ret

Assembly code 4: _cdecl calling convention

3) _fastcall

_fastcall stores the parameters in the registers ECX and EDX. This type of calling

convention is used to increase the performance of the function, as register access is faster

than stack access. This convention also does not have any function returns.

011411D0 push ebp

011411D1 mov ebp,esp

011411D3 sub esp,48h

011411D6 push ebx

011411D7 push esi

011411D8 push edi

011411D9 mov dword ptr [ebp-8],edx

011411DC mov dword ptr [ebp-4],ecx

011411DF mov eax,dword ptr [ebp-4]

011411E2 cmp eax,dword ptr [ebp-8]

011411E5 jge 011411F5

011411E7 push 1145720h

011411EC call dword ptr ds:[1148248h]

011411F2 add esp,4

011411F5 pop edi

011411F6 pop esi

011411F7 pop ebx

011411F8 mov esp,ebp

011411FA pop ebp

011411FB ret

Assembly code 5: _fastcall calling convention

5.2 Algorithm

Figure 4 depicts the flowchart of the decompiler algorithm explained in this paper. All three

different passes and their sub-modules are explained in further sections of this chapter. The

background information required to understand this algorithm has been provided in previous

chapters. Each of these passes is explained with one test case.

23

Figure 4: Decompiling algorithm flowchart

24

5.1 First Pass – Raw view

This pass deals with reading the assembly language from the .dis file and storing it in a vector of

strings. Each string is one assembly instruction read from the file in the specific format of

memory address, assembly instruction, operand 1, and operand 2. Each instruction follows this

specific format and order, and hence can easily be read using regular expressions. Instruction

prefixes and three-operand instructions are currently unsupported. Operand 1 and operand 2 are

separated by a comma delimiter, and operand 2 is optional for some instructions. This pass

detects if the input assembly file contains any errors. If so, the algorithm is terminated and a

corresponding error message is displayed.

5.1.1 Raw view example

This section describes an example of the complete run of this pass along with the format in

which the assembly instruction is read.

Consider the following input assembly code:

011411A0 push ebp

011411A1 mov ebp,esp

011411A3 sub esp,40h

011411A6 push ebx

011411A7 push esi

011411A8 push edi

011411A9 mov eax,dword ptr [ebp+8]

011411AC cmp eax,dword ptr [ebp+0Ch]

011411AF jge 011411BF

011411B1 push 1145720h

011411B6 call dword ptr ds:[1148248h]

011411BC add esp,4

011411BF pop edi

011411C0 pop esi

011411C1 pop ebx

011411C2 mov esp,ebp

011411C4 pop ebp

011411C5 ret

Figure 5: Raw view input

The algorithm reads the assembly code line-by-line to the end of file, in order to generate a

vector of strings in the format shown below:

Table 4: Assembly instruction string format

Memory address Instruction Operand 1 Operand 2

011411A9 mov eax dword ptr [ebp + 8]

25

5.2 Second Pass – Parsed view

This pass deals with generating an intermediate representation from the raw view for the

algorithm, which is very easy to handle and manipulate by logic during the recognized view. The

output of raw view consists of a vector of strings that is very difficult to handle in programming

logic because it involves string manipulation functions. This step takes the vector of strings as

input and generates an IR better suited for the algorithm to understand the input assembly and to

generate the pseudo C output. This step has no major role in the core logic of the decompilation

process, but it makes the algorithm more robust and increases overall performance.

This step performs three major operations on the assembly code from the raw view. They are

explained below in the order they are performed.

Conversion of vector strings to structures

The raw input consists of strings corresponding to each assembly instruction from the initial

input. These strings contain the most important details that will be used by the recognized view.

However, handling the strings in a complicated algorithm can be very tedious and non-

optimized programming; so, this step tries to solve this problem by converting each string into a

C structure that can then be used in the logic very efficiently. The structure of this C struct is as

follows:

typedef struct _RECOGNIZED_INSTRUCTION

{

ULONG_PTR AbsoluteAddress;

ULONG_PTR RelativeAddress;

UINT InstructionId;

OPERAND Operand1;

OPERAND Operand2;

} RECOGNIZED_INSTRUCTION, *PRECOGNIZED_INSTRUCTION;

Code 8: Recognized instruction C structure

Relative Addressing

From the previous section, we have seen that an assembly instruction contains a memory

location. This memory location is the actual location of that instruction when the executable is

loaded into memory. These locations are of critical importance to the algorithm because the

26

assembly language processes loops and execution flows based on these addresses. The memory

location of the successive instruction is added to the instruction size of the previous instruction;

so, it can be derived that the memory locations are successive in memory. Their initial starting

point depends on the operating system, and is likely to change on every execution run. But,

since this algorithm focuses on decompilation on a smaller scale, those factors are not of much

concern.

The locations are eight digits long, but the first half will likely be the same for each instruction

unless the executable is very large. Again, this is constrained in our algorithm, so the important

parts from the location for our algorithm are the lower-order bits. Since the input assembly is not

the complete assembly of the executable, but only the section of it that we are interested in

decompiling, the actual address can be replaced by the relative address as long as we preserve

the instruction size.

Instruction ID

The algorithm covers a specific set of instructions from the Intel x86 instruction manual [12].

This step assigns a constant value to each of the instructions in the raw view. Assigning an

integer value to the string helps us to identify the instruction in the complex logic through

conditions rather than string comparisons. The raw view pass has already detected if the code

consists of unsupported instructions – the execution reaches this pass only with supported

instructions. The instructions supported by the algorithm are as follows:

Table 5: Supported assembly instructions

Functionality Instructions

Math Operations

Fld, Fadd, Fstp, Fsub, Fmul, Fdiv, Imul,

Add, Sub

Stack Operations Push, Pop

Function Operations Ret, Call

Jump Operations Ja, Je, Jle, Jge, Jne, Jmp

General-Purpose Operations Cmp, Mov

27

5.2.1 Parsed view example

This example shows the changes made by parsed view to the input assembly code:

Input assembly code to the parsed view

011411A0 push ebp

011411A1 mov ebp,esp

011411A3 sub esp,40h

011411A6 push ebx

011411A7 push esi

011411A8 push edi

011411A9 mov eax,dword ptr [ebp+8]

011411AC cmp eax,dword ptr [ebp+0Ch]

011411AF jge 011411BF

011411B1 push 1145720h

011411B6 call dword ptr ds:[1148248h]

011411BC add esp,4

011411BF pop edi

011411C0 pop esi

011411C1 pop ebx

011411C2 mov esp,ebp

011411C4 pop ebp

011411C5 ret

Assembly code 6: Parsed view input

Output generated by parsed view

Figure 6: Parsed view output

5.3 Third Pass – Recognized view

This pass contains the core logic of the algorithm and deals with parsing the parsed view and

creating pseudo C code. This process can be broken down into a number of steps, each

28

performing different specific operations from the input received from the previous step. Each

decompilation process in the algorithm follows these steps:

1) Stack initialization and analysis

2) Register initialization and analysis

3) Graph generation

4) Output generation

5.3.1 Stack initialization and analysis

Each function call in the assembly language generates a few stack instructions before executing

the function logic. These stack instructions contain important information – calling convention

used, number of parameters, and their values and function return location. This step analyzes

these instructions and captures the most important information needed for the further steps in the

algorithm.

The initialization of the stack structure helps us to determine our calling convention and

parameter information.

Figure 7: Stack initialization

29

The use of ESP determines whether the calling convention used was _cdecl or _stdcall, but it

cannot still be determined as differentiating them based on the RET instruction. As in the _cdecl

calling convention, the caller cleans up the stack after a successful function call; in _stdcall, the

callee pops the parameters from the stack. Also, each parameter passed to the stack takes 4 bytes

of memory space. This helps us determine the number of parameters used in the underlying

logic. The space below the stack pointer helps us determine the number of local variables created

by the function.

5.3.2 Register initialization and analysis

Intel x86 uses eight general purpose registers for executing assembly instructions. These

registers are used by the processor for an immediate reference, as accessing these registers is

faster than physical memory. The analysis of these registers helps determine the _fastcall calling

convention and the intermediate values used while performing the calculation. The _fastcall

calling convention uses ECX and EDX registers for storing the parameters passed to the

function. EAX and EBX are used for immediate calculation and storing intermediate values [12].

We recorded the states of each register during the course of the algorithm whenever the assembly

instruction refers to keywords related to registers. The initialization of the registers is as follows:

Figure 8: Register initialization

30

ECX and EDX help determine the parameters passed to the function and the calling convention

used, while EAX and EBX are useful during analysis of the loop counters and other immediate

variables used in the assembly instruction.

5.3.3 Graph Generation

This is the most critical and essential part of the algorithm, determining the high-level

programming language syntax from the underlying assembly language. This step reads the most

steps of the raw view at a single time compared to any other steps in the algorithm, and hence is

the most important step of this research paper.

Each supported assembly instruction has a function handler, which contains the logic on how to

handle the specific assembly instruction. Some assembly instructions are standalone, and are

most equivalent to their respective C instructions in terms of granularity – for example, MOV.

However, most assembly instructions are highly dependent on further instructions to provide

simple instructions in high-level programming language. Each of these function handlers

contains the logic on how to handle its own assembly instruction, and each handler sets the flags

accordingly for the further steps of the algorithm.

Based on the information returned from these function handlers, the next process helps us

determine the more complex C syntaxes, like loops. This step generates dependency graphs and

semantic loads by virtually executing each assembly instruction [10]. The semantic load helps us

to determine the loop counters, and dependency graphs determine the type of structure of

execution, like for loop, switch, or if/else. We do not differentiate between for loops and while

loops, since that information is lost during compilation – all for loops can be represented by

while loops, so this is not a major constraint to this step. Figure 9 shows the different function

handlers implemented into this algorithm. The algorithm is designed in this manner to allow easy

implementation of support for more instructions in the future. Any instruction to be supported in

this algorithm must have a function handler defined in this section of code, and the logic to be

implemented in its function handler.

31

static INSTRUCTION SupportedInstructions[] =

{

{ L"add", AddHandler },

{ L"call", CallHandler },

{ L"cmp", CmpHandler },

{ L"ja", JaHandler },

{ L"je", JeHandler },

{ L"jge", JgeHandler },

{ L"jle", JleHandler },

{ L"jne", JneHandler },

{ L"jmp", JmpHandler },

{ L"imul", ImulHandler },

{ L"mov", MoveHandler },

{ L"pop", PopHandler },

{ L"push", PushHandler },

{ L"ret", RetHandler },

{ L"sub", SubHandler },

{ L"fld", FldHandler },

{ L"fadd", FaddHandler },

{ L"fstp", FstpHandler },

{ L"fsub", FsubHandler },

{ L"fmul", FmulHandler },

{ L"fdiv", FdivHandler }

};

Code 9: Supported instructions' function handlers

Every assembly instruction is treated as a vertex of the graph, and execution flow can be

considered an edge to build this dependency graph [21]. The beginning and end of the function

are also considered vertexes. We do not consider the stack and register operations, as they are

already handled by the previous steps of this pass. We use this graph to build the incidence list.

Below is the complete example of original C code with nested if along with the respective

dependency graph and incidence list and its interpretation:

Original C code

If(a > 4){

If( a != 5){

Foo();

}

}Return;

Code 10: Nested if

32

Dependency graph for nested if

Figure 9: Nested if dependency graph

This graph can help us build the incidence lists. Incidence lists are very convenient for

recognizing high-level programming structures [22]. The incidence list for the above graph

would be,

(1, 2)

(2, 3), (2, 5)

(3, 4), (3, 5)

(4, 5)

Interpretation of incidence list

(2, 3), (2, 5) represents if condition a>5

(3, 4), (3, 5) represents the nested if condition a != 5

(4, 5) represents the end of the if condition with both if clauses satisfied

33

If-then detection

If the case is only if-then, then we have a different graph compared to mixed if-then and nested

ifs. In order to identify if-then, we have to detect the following graph pattern and ensure that

there is no external reference to (x+1).

Dependency graph

Figure 10: If-then dependency graph

Incidence list

(x, x+1)

(x, x+2)

(x+1, x+2)

Interpretation of incidence list

(x, x+1) represents inside the if condition

(x, x+2) represents outside the else condition

(x+1, x+2) represents the end of the if condition with the if clause dissatisfied

34

If/else detection

Detecting if/else is very similar to previous cases, except for the difference in the incidence list.

If we have an if/else inside an if, that makes a compound if statement and we can replace two

if/elses with one condition. We must do that to eliminate all ifs.

Dependency Graph

Incidence list

(x, x+1)

(x, x+2)

(x+1, x+3)

(x+2, x+3)

Interpretation of incidence list

(x, x+1) represents inside the if condition

(x, x+2) represents inside the else condition

(x+1, x+3) represents the end of the if/else with the if condition satisfied

(x+2, x+3) represents the end of the if/else with the else condition satisfied

Figure

11

: Dependency graph for if/else

35

Loop detection

Loops primarily create three graphs based on three types of loops: the normal loop, continue

loop, and break loop. Normal loops consist of top-to-down direction graphs, but the edge from

the last node leads to the start node. The start node also has an edge leading to finish node based

upon the condition. This case is similar in all type of loops.

Continue loops have an edge going from the middle nodes to the start node, while break loops

have an edge going from the middle nodes to the outside of the loop nodes. Figure 12 depicts a

for loop with an if/else statement that contains the switch and break cases. This helps us to

determine the difference in detecting these loops.

Factors like conditions and loop counters have already been handled by the function handlers in

the previous step.

Dependency graph

Figure 12: Dependency graph for loop

36

Incidence list

(1, 2)

(2, 3), (2, 7)

(3, 4), (3, 5)

(4, 5), (4, 2)

(5, 6), (5, 7)

(6, 2)

Interpretation of the incidence list

(2, 3), (2, 7) represents the for loop condition

(3, 4), (3, 5) represents the if/else condition

(4, 5), (4, 2) represents continue

(5, 6), (5, 7) represents break

(6, 2) represents one loop iteration

Switch detection

Switch case statements are similar to if/then statements. However, if the code consists of

multiple ifs accessing the same variable, it is very difficult to determine the difference between

them as they will produce the same dependency graph. The switch is detected if the graph

consists of a vertex with many outgoing edges to different nodes, which then have outgoing

edges to a finish node. The default case can be detected, as there will be an edge with no

condition. Figure 13 on the next page depicts the dependency graph.

37

Dependency graph

Incidence list

(1, 2)

(2, 3), (2, 4), (2, 5), (2, 6)

(3, 6)

(4, 6)

(5, 6)

Interpretation of the incidence list

(1, 2) represents the switch condition

(2, 3), (2, 4), (2, 5), (2, 6) represents different cases

(3, 6) represents the end of the switch case with case 1 satisfied

(4, 6) represents the end of the switch case with case 1 satisfied

(5, 6) represents the end of the switch case with case 1 satisfied

Figure 13: Dependency graph for switch case

38

5.3.4 Output generation

This step produces the final output of the algorithm: the pseudo C code. The major functions of

this step include variable naming, syntax structuring, and filling the information gaps left over

from the previous steps. This step does not involve much complicated logic, as much of the

information necessary to complete this step is available from the previous steps.

Variable naming

The register-analyzing and stack-analyzing steps help us to determine the number of parameters

passed to this function or the number of local variables created. Retrieving the original variable

names is not possible in this process, so we assign new variable names to our newly-found

variables. The naming convention names them according to “ParamXX,” where XX stands for

increasing order – 00, 01, 02.

Function definition

We obtained the information about the function calling convention and parameters from the

previous steps. Determining the function return type is currently not within the scope of this

project. This step creates the function prototype based on the information we have acquired.

Syntax structuring

The graph generation phase has already provided the high-level programming logic of the

reversed assembly code. Now, this step handles the parenthesis “{“ and “}” around this code,

relying on regular expressions and flags from the previous steps.

Gap filling

Some assembly codes that refer to a different memory location outside of input through the use

of CALL functions are not possible to decompile with this algorithm. Library or system function

references are examples of this – they are represented in the output by their same assembly

instruction.

39

5.3.5 Output

Below is the final output created by the algorithm. This example consists of three parameter

inputs, for loop, switch case, and if-then-else. The output has the look and feel of a C code, and

is more understandable than related assembly code. Printf statements are represented by call

instructions to the related address in data section.

Figure 14: Recognized view output

40

6. Test cases and Results

In this section we discuss a complete test case along with its results. More test cases with

different original code syntaxes and structures are provided in Appendix A. This particular test

case consists of for loop, switch case, and if/else in the original code. The compiler used to

compile test cases is Microsoft Visual C++, and the assembler is Microsoft x86 Assembler.

Optimization has been disabled for generating the test cases. The original C code, generated

assembly code, and decompiler-generated output is provided below:

Original Code

void __fastcall Test14b ( int a, int b, int c )

{

for ( int i = 0; i < c; i++ )

{

if ( i == a )

break;

printf ( "loop" );

}

switch ( a )

{

case 1:

printf ( "a = 1\n" );

break;

case 2:

printf ( "a = 2\n" );

break;

}

printf ( "in the middle\n" );

if ( a > b )

printf ( "a > b\n" );

else

printf ( "a <= b\n" );

}

Code 11: Test case C code

41

Assembly Code

012D2320 push ebp

012D2321 mov ebp,esp

012D2323 sub esp,50h

012D2326 push ebx

012D2327 push esi

012D2328 push edi

012D2329 mov dword ptr [ebp-8],edx

012D232C mov dword ptr [ebp-4],ecx

012D232F mov dword ptr [ebp-0Ch],0

012D2336 jmp 012D2341

012D2338 mov eax,dword ptr [ebp-0Ch]

012D233B add eax,1

012D233E mov dword ptr [ebp-0Ch],eax

012D2341 mov eax,dword ptr [ebp-0Ch]

012D2344 cmp eax,dword ptr [ebp+8]

012D2347 jge 012D2363

012D2349 mov eax,dword ptr [ebp-0Ch]

012D234C cmp eax,dword ptr [ebp-4]

012D234F jne 012D2353

012D2351 jmp 012D2363

012D2353 push 12D5870h

012D2358 call dword ptr ds:[12D8248h]

012D235E add esp,4

012D2361 jmp 012D2338

012D2363 mov eax,dword ptr [ebp-4]

012D2366 mov dword ptr [ebp-50h],eax

012D2369 cmp dword ptr [ebp-50h],1

012D236D je 012D2377

012D236F cmp dword ptr [ebp-50h],2

012D2373 je 012D2387

012D2375 jmp 012D2395

012D2377 push 12D5854h

012D237C call dword ptr ds:[12D8248h]

012D2382 add esp,4

012D2385 jmp 012D2395

012D2387 push 12D5840h

012D238C call dword ptr ds:[12D8248h]

012D2392 add esp,4

012D2395 push 12D5880h

012D239A call dword ptr ds:[12D8248h]

012D23A0 add esp,4

012D23A3 mov eax,dword ptr [ebp-4]

012D23A6 cmp eax,dword ptr [ebp-8]

012D23A9 jle 012D23BB

012D23AB push 12D5734h

012D23B0 call dword ptr ds:[12D8248h]

012D23B6 add esp,4

012D23B9 jmp 012D23C9

012D23BB push 12D5784h

012D23C0 call dword ptr ds:[12D8248h]

012D23C6 add esp,4

012D23C9 pop edi

012D23CA pop esi

012D23CB pop ebx

012D23CC mov esp,ebp

012D23CE pop ebp

012D23CF ret 4

Assembly code 7: Test case assembly code

42

Decompiler-Generated Output

6.1 Comparison with manual decompilation

This test case involves manually decompiling assembly code and comparing it with the

algorithm-generated output. The assembly code was provided to a user with some assembly

language knowledge, who was then asked to generate a C code. The test case chosen was quite

simple, containing only an if/else. The original C code, assembly code, manually-generated code,

and decompiler-generated output are provided below:

Figure 15: Decompiler-generated output

43

Original Code

void __cdecl Test2a ( int a, int b )

{

if ( a < b )

printf ( "a < b\n" );

else

printf ( "a >= b\n" );

}

Code 12: if/else original C code

Assembly Code

01141250 push ebp

01141251 mov ebp,esp

01141253 sub esp,40h

01141256 push ebx

01141257 push esi

01141258 push edi

01141259 mov eax,dword ptr [ebp+8]

0114125C cmp eax,dword ptr [ebp+0Ch]

0114125F jge 01141271

01141261 push 1145720h

01141266 call dword ptr ds:[1148248h]

0114126C add esp,4

0114126F jmp 0114127F

01141271 push 1145728h

01141276 call dword ptr ds:[1148248h]

0114127C add esp,4

0114127F pop edi

01141280 pop esi

01141281 pop ebx

01141282 mov esp,ebp

01141284 pop ebp

01141285 ret

Assembly code 8: Assembly code if/else

Manually-Generated C Code

void test ( )

{

int i, j;

if ( i < j )

do something;

if ( i > j )

do something;

}

Code 13: Manually-generated output

44

Decompiler-Generated C Code

Figure 16: Decompiler-generated output

Analysis

As we can see, the user was unable to determine the calling convention used for the function call.

As a result, the user was unable to find the number of parameters passed to the algorithm, and

this leads to the assumption of two local variables i and j. Also, the original syntax of if/else was

not replaced by two simple ifs, and the user was unable to determine the details inside the if

statements.

45

8. Future Work

From this research, we can conclude that decompilation is a very complicated process, and is

constrained by a number of limitations. The proposed algorithm tries to implement some of the

concepts in the above-mentioned process in a very effective way. However, many improvements

could still be added to this algorithm for a more successful reversing process and to cover more

assembly instructions.

More complicated assembly instructions

If other assembly instructions that are more complicated and not covered in this paper could be

added to this algorithm, it could improve the reversing of highly-complicated code. The

algorithm is adaptable to the addition of other instructions by adding them to the function

mapping and implementing the corresponding functionality of that instruction in its relative

function. If the assembly instruction to be added is more dependent on execution flow of the

code, then it would require a considerable amount of testing with the assembly to determine the

changes in dependency graphs created by these instructions [7].

Instructions that are standalone or that do not affect the execution flow can easily be

implemented with the above approach. Before adding the implementation of the assembly

instructions, the instruction’s impact on the output code needs to be studied deeply. Some

assembly instructions are created specifically for hardware operations that do not have

corresponding high-level language implementations. Decompilation algorithms are heavily based

on assumptions [1] [3], but we would like to keep the assumption factor of the algorithm as low

as possible. One possible way to achieve this is by not adding extra support for instructions that

do not play a major role in the related output.

Handling the data section with the code section

This algorithm is currently limited to handling only the .code section of the assembly language.

Future work could include modifying the algorithm to handle both the .data section and the .code

section [12]. This could help to achieve the constants and values used in high-level language.

46

This feature can retrieve the strings and other data that gets separated from the code section

during the compilation process [15].

Inbuilt disassembler

The current scenario takes the .dis file as an input and generates a text C file. If a further step to

disassemble – that is, reversing the machine code to the assembly code from an executable file –

is added to the algorithm, we could avoid the manual file creation of the .dis file. Also, the

algorithm would have more information, along with all the sections that it currently lacks.

Object oriented code and user interface

These two topics have been the least touched on in this process, as reversing them is extremely

complicated and mostly unreliable in terms of the output generated. The assembly code removes

the concept of the object oriented code when it is compiled [7], and the UI code is handled by the

system calls. With different compiler settings and a controlled environment, we could still

attempt to overcome these barriers.

47

Appendix A – Test cases and results

If /else Statements

Original Code

void __cdecl Test2a ( int a, int b )

{

if ( a < b )

printf ( "a < b\n" );

else

printf ( "a >= b\n" );

}

Assembly Code

01141250 push ebp

01141251 mov ebp,esp

01141253 sub esp,40h

01141256 push ebx

01141257 push esi

01141258 push edi

01141259 mov eax,dword ptr [ebp+8]

0114125C cmp eax,dword ptr [ebp+0Ch]

0114125F jge 01141271

01141261 push 1145720h

01141266 call dword ptr ds:[1148248h]

0114126C add esp,4

0114126F jmp 0114127F

01141271 push 1145728h

01141276 call dword ptr ds:[1148248h]

0114127C add esp,4

0114127F pop edi

01141280 pop esi

01141281 pop ebx

01141282 mov esp,ebp

01141284 pop ebp

01141285 ret

Decompiled Output

48

Nested if /else statements

Original Code

void __cdecl Test3a ( int a, int b )

{

if ( a < b )

printf ( "a < b\n" );

else

if ( a == b )

printf ( "a = b\n" );

else

printf ( "a > b\n" );

}

Assembly Code

01141340 push ebp

01141341 mov ebp,esp

01141343 sub esp,40h

01141346 push ebx

01141347 push esi

01141348 push edi

01141349 mov eax,dword ptr [ebp+8]

0114134C cmp eax,dword ptr [ebp+0Ch]

0114134F jge 01141361

01141351 push 1145720h

01141356 call dword ptr ds:[1148248h]

0114135C add esp,4

0114135F jmp 01141387

01141361 mov eax,dword ptr [ebp+8]

01141364 cmp eax,dword ptr [ebp+0Ch]

01141367 jne 01141379

01141369 push 114573Ch

0114136E call dword ptr ds:[1148248h]

01141374 add esp,4

01141377 jmp 01141387

01141379 push 1145734h

0114137E call dword ptr ds:[1148248h]

01141384 add esp,4

01141387 pop edi

01141388 pop esi

01141389 pop ebx

0114138A mov esp,ebp

0114138C pop ebp

0114138D ret

Decompiled Output

49

Complex if /else statements

Original Code

void __fastcall Test4b ( int a, int b, int c )

{

if ( a == b && b == c )

printf ( "a = b = c\n" );

}

Assembly Code

011414D0 push ebp

011414D1 mov ebp,esp

011414D3 sub esp,48h

011414D6 push ebx

011414D7 push esi

011414D8 push edi

011414D9 mov dword ptr [ebp-8],edx

011414DC mov dword ptr [ebp-4],ecx

011414DF mov eax,dword ptr [ebp-4]

011414E2 cmp eax,dword ptr [ebp-8]

011414E5 jne 011414FD

011414E7 mov eax,dword ptr [ebp-8]

011414EA cmp eax,dword ptr [ebp+8]

011414ED jne 011414FD

011414EF push 1145744h

011414F4 call dword ptr ds:[1148248h]

011414FA add esp,4

011414FD pop edi

011414FE pop esi

011414FF pop ebx

01141500 mov esp,ebp

01141502 pop ebp

01141503 ret 4

Decompiled Output

50

Complex if /else statements

Original Code

void __stdcall Test6c ( int a, int b, int c )

{

if ( a == b && b != c && a > c )

printf ( "a = b and b != c and a > c\n" );

}

Assembly Code

013C1720 push ebp

013C1721 mov ebp,esp

013C1723 sub esp,40h

013C1726 push ebx

013C1727 push esi

013C1728 push edi

013C1729 mov eax,dword ptr [ebp+8]

013C172C cmp eax,dword ptr [ebp+0Ch]

013C172F jne 013C174F

013C1731 mov eax,dword ptr [ebp+0Ch]

013C1734 cmp eax,dword ptr [ebp+10h]

013C1737 je 013C174F

013C1739 mov eax,dword ptr [ebp+8]

013C173C cmp eax,dword ptr [ebp+10h]

013C173F jle 013C174F

013C1741 push 13C5768h

013C1746 call dword ptr ds:[13C8248h]

013C174C add esp,4

013C174F pop edi

013C1750 pop esi

013C1751 pop ebx

013C1752 mov esp,ebp

013C1754 pop ebp

013C1755 ret 0Ch

Decompiled Output

51

Complex nested if/else ladder

Original Code

void __cdecl Test7a ( int a, int b, int c )

{

if ( a < b )

{

if ( b < c )

printf ( "a < b and b < c\n" );

else if ( b == c )

printf ( "a < b and b = c\n" );

else

printf ( "a < b and b > c\n" );

} else

if ( a == b )

{

if ( b < c )

printf ( "a = b and b < c\n" );

else if ( b == c )

printf ( "a = b and b = c\n" );

else

printf ( "a = b and b > c\n" );

} else

{

if ( b < c )

printf ( "a > b and b < c\n" );

else if ( b == c )

printf ( "a > b and b = c\n" );

else

printf ( "a > b and b > c\n" );

}

52

Assembly Code

01141720 push ebp

01141721 mov ebp,esp

01141723 sub esp,40h

01141726 push ebx

01141727 push esi

01141728 push edi

01141729 mov eax,dword ptr [ebp+8]

0114172C cmp eax,dword ptr [ebp+0Ch]

0114172F jge 01141774

01141731 mov eax,dword ptr [ebp+0Ch]

01141734 cmp eax,dword ptr [ebp+10h]

01141737 jge 01141749

0114173E call dword ptr ds:[1148248h]

01141744 add esp,4

01141747 jmp 0114176F

01141749 mov eax,dword ptr [ebp+0Ch]

0114174C cmp eax,dword ptr [ebp+10h]

0114174F jne 01141761

01141751 push 1145818h

01141756 call dword ptr ds:[1148248h]

0114175C add esp,4

0114175F jmp 0114176F

01141761 push 1145804h

01141766 call dword ptr ds:[1148248h]

0114176C add esp,4

01141774 mov eax,dword ptr [ebp+8]

01141777 cmp eax,dword ptr [ebp+0Ch]

0114177A jne 011417BC

0114177C mov eax,dword ptr [ebp+0Ch]

0114177F cmp eax,dword ptr [ebp+10h]

01141782 jge 01141794

01141784 push 11457F0h

01141789 call dword ptr ds:[1148248h]

0114178F add esp,4

01141792 jmp 011417BA

01141794 mov eax,dword ptr [ebp+0Ch]

01141797 cmp eax,dword ptr [ebp+10h]

0114179A jne 011417AC

0114179C push 11457DCh

011417A1 call dword ptr ds:[1148248h]

011417A7 add esp,4

011417AA jmp 011417BA

011417AC push 11457C8h

011417B7 add esp,4

011417BA jmp 011417FA

011417BC mov eax,dword ptr [ebp+0Ch]

011417BF cmp eax,dword ptr [ebp+10h]

011417C2 jge 011417D4

011417C4 push 11457B4h

011417CF add esp,4

011417D2 jmp 011417FA

011417D4 mov eax,dword ptr [ebp+0Ch]

011417D7 cmp eax,dword ptr [ebp+10h]

011417DA jne 011417EC

011417DC push 11457A0h

011417E1 call dword ptr ds:[1148248h]

011417E7 add esp,4

011417EA jmp 011417FA

011417EC push 114578Ch

011417F1 call dword ptr ds:[1148248h]

011417F7 add esp,4

011417FA pop edi

011417FB pop esi

011417FC pop ebx

011417FD mov esp,ebp

011417FF pop ebp

01141800 ret

53

Decompiled Output

54

Switch case

Original Code

void __stdcall Test9c ( int a )

{

switch ( a )

{

case 0:

printf ( "a = 0\n" );

break;

case 1:

printf ( "a = 1\n" );

break;

case 2:

printf ( "a = 2\n" );

break;

default:

printf ( "a out of range\n" );

break;

}

Assembly Code

00BF1E50 push ebp

00BF1E51 mov ebp,esp

00BF1E53 sub esp,44h

00BF1E56 push ebx

00BF1E57 push esi

00BF1E58 push edi

00BF1E59 mov eax,dword ptr [ebp+8]

00BF1E5C mov dword ptr [ebp-44h],eax

00BF1E5F cmp dword ptr [ebp-44h],0

00BF1E63 je 00BF1E73

00BF1E65 cmp dword ptr [ebp-44h],1

00BF1E69 je 00BF1E83

00BF1E6B cmp dword ptr [ebp-44h],2

00BF1E6F je 00BF1E93

00BF1E71 jmp 00BF1EA3

00BF1E73 push 0BF585Ch

00BF1E78 call dword ptr ds:[0BF8248h]

00BF1E7E add esp,4

00BF1E81 jmp 00BF1EB1

00BF1E83 push 0BF5854h

00BF1E88 call dword ptr ds:[0BF8248h]

00BF1E8E add esp,4

00BF1E91 jmp 00BF1EB1

00BF1E93 push 0BF5840h

00BF1E98 call dword ptr ds:[0BF8248h]

00BF1E9E add esp,4

00BF1EA1 jmp 00BF1EB1

55

00BF1EA3 push 0BF5BA8h

00BF1EA8 call dword ptr ds:[0BF8248h]

00BF1EAE add esp,4

00BF1EB1 pop edi

00BF1EB2 pop esi

00BF1EB3 pop ebx

00BF1EB4 mov esp,ebp

00BF1EB6 pop ebp

00BF1EB7 ret 4

Decompiled Output

For loop with break and continue

Original Code

void __stdcall Test13c ( int a, int b, int c )

{

for ( int i = 0; i < c; i += 3 )

{

if ( i == b )

break;

if ( i == a )

continue;

printf ( "%d\n", i );

}

56

Assembly Code

012D1EC0 push ebp

012D1EC1 mov ebp,esp

012D1EC3 sub esp,44h

012D1EC6 push ebx

012D1EC7 push esi

012D1EC8 push edi

012D1EC9 mov dword ptr [ebp-4],0

012D1ED0 jmp 012D1EDB

012D1ED2 mov eax,dword ptr [ebp-4]

012D1ED5 add eax,3

012D1ED8 mov dword ptr [ebp-4],eax

012D1EDB mov eax,dword ptr [ebp-4]

012D1EDE cmp eax,dword ptr [ebp+10h]

012D1EE1 jge 012D1F0B

012D1EE3 mov eax,dword ptr [ebp-4]

012D1EE6 cmp eax,dword ptr [ebp+0Ch]

012D1EE9 jne 012D1EED

012D1EEB jmp 012D1F0B

012D1EED mov eax,dword ptr [ebp-4]

012D1EF0 cmp eax,dword ptr [ebp+8]

012D1EF3 jne 012D1EF7

012D1EF5 jmp 012D1ED2

012D1EF7 mov eax,dword ptr [ebp-4]

012D1EFA push eax

012D1EFB push 12D586Ch

012D1F00 call dword ptr ds:[12D8248h]

012D1F06 add esp,8

012D1F09 jmp 012D1ED2

012D1F0B pop edi

012D1F0C pop esi

012D1F0D pop ebx

012D1F0E mov esp,ebp

012D1F10 pop ebp

012D1F11 ret 0Ch

Decompiled Output

57

Floating point math operations

Original Code

void Test15 ( float a, float b )

{

float c;

c = a + b;

printf ( "%f\n", c );

c = a - b;

printf ( "%f\n", c );

c = a * b;

printf ( "%f\n", c );

c = a / b;

printf ( "%f\n", c );

}

Assembly Code

012D2490 push ebp

012D2491 mov ebp,esp

012D2493 sub esp,44h

012D2496 push ebx

012D2497 push esi

012D2498 push edi

012D2499 fld dword ptr [ebp+8]

012D249C fadd dword ptr [ebp+0Ch]

012D249F fstp dword ptr [ebp-4]

012D24A2 fld dword ptr [ebp-4]

012D24A5 sub esp,8

012D24A8 fstp qword ptr [esp]

012D24AB push 12D5730h

012D24B0 call dword ptr ds:[12D8248h]

012D24B6 add esp,0Ch

012D24B9 fld dword ptr [ebp+8]

012D24BC fsub dword ptr [ebp+0Ch]

012D24BF fstp dword ptr [ebp-4]

012D24C2 fld dword ptr [ebp-4]

012D24C5 sub esp,8

012D24C8 fstp qword ptr [esp]

012D24CB push 12D5730h

012D24D0 call dword ptr ds:[12D8248h]

012D24D6 add esp,0Ch

012D24D9 fld dword ptr [ebp+8]

012D24DC fmul dword ptr [ebp+0Ch]

012D24DF fstp dword ptr [ebp-4]

012D24E2 fld dword ptr [ebp-4]

012D24E5 sub esp,8

012D24E8 fstp qword ptr [esp]

012D24EB push 12D5730h

012D24F0 call dword ptr ds:[12D8248h]

012D24F6 add esp,0Ch

012D24F9 fld dword ptr [ebp+8]

012D24FC fdiv dword ptr [ebp+0Ch]

012D24FF fstp dword ptr [ebp-4]

012D2502 fld dword ptr [ebp-4]

012D2505 sub esp,8

012D2508 fstp qword ptr [esp]

58

012D250B push 12D5730h

012D2510 call dword ptr ds:[12D8248h]

012D2516 add esp,0Ch

012D2519 pop edi

012D251A pop esi

012D251B pop ebx

012D251C mov esp,ebp

012D251E pop ebp

012D251F ret

Decompiled Output

59

References

1. M. Emmerich (2007). Static Single Assignment for Decompilation.

2. P. Enzo (2009). ASM to C Translation table.

3. Decompiler topics (2011) retrieved from:

http://www.compwisdom.com/topics/Decompiler

4. Decompilation Wiki (2011) retrieved from:

http://program-transformation.org/Transform/DeCompilation

5. The dcc compiler (2011) retrieved from:

http://www.itee.uq.edu.au/~cristina/dcc.html

6. REC compiler (2011) retrieved from:

http://www.backerstreet.com/rec/rec.htm

7. Paul Carter (2004). PC assembly language.

8. Intel assets (2011). Intel 64 and IA32 architecture. Volume 2 (2A & 2B): Instruction Set

Reference, A-Z.

9. E. Chikofsky and J. Cross (1990). Reverse engineering and design recovery.

10. C. Mills, S. Ahalt, and J. Fowler (1991). Compiled instruction set simulation. Software _

Practice and Experience.

11. G. Morrisett, D. Walker, K. Crary, and N. Glew (1998). From System F to typed

assembly language.

12. Matt Pietrek (2002). An In-Depth Look into the Win32 Portable Executable File Format.

February 2002 issue of MSDN Magazine.

13. C. Cifuentes. (1994). Reverse Compilation Techniques.

14. G. Balakrishnan, T. Reps, D. Melski, and T. Teitelbaum (2005). WYSINWYX:What You

See Is Not What You eXecute.

15. R. Horspool and N. Marovac (1979). An approach to the problem of detranslation of

computer programs.

16. D. Eriksson (2002). Desquirr web page.

17. A. Mycroft (1999). Type-based decompilation.

18. Andromeda decompiler (2004). Andromeda decompiler web page.

19. I. Guilfanov (2007). Decompilation gets real.

60

20. Austin Code Works (1990), Exec-2 Decompiler

21. Balmas, Francoise (2001).

Displaying dependence graphs: a hierarchical approach.

22. M. T. Goodrich and R. Tamassia (1998). Data Structures and Algorithms.

23. Scott, M. Lee (2005). Programming Language Pragmatics.

24. G. Caprino (2003). REC - Reverse Engineering Compiler.

25. S. Kumar (2001). Disc.

26. J. Miecznikowski and L. Hendren (2002). Decompiling Java bytecode: Problems,

traps and pitfalls.

27. J. Hoenicke (2000). Java Optimize and Decompile Environment.

28. P. Kouznetsov (1999). JAD - the fast JAva Decompiler.

29. K. Kumar (2001). JReversePro - Java decompiler.

30. V. Schneider and G. Winiger (1974). Translation grammars for compilation and

Decompilation.

31. J. Reuter (1988). Public domain software.

32. Muchnick, Steven (1997). Advanced Compiler Design and Implementation.