Reversing Microsoft Visual C++ Part I: Exception Handling

Abstract

Microsoft Visual C++ is the most widely used compiler for Win32 so it is important for the Win32 reverser to be familiar with its inner working. Being able to recognize the compiler-generated glue code helps to quickly concentrate on the actual code written by the programmer. It also helps in recovering the high-level structure of the program.

In part I of this 2-part article (see also: Part II: Classes, Methods and RTTI), I will concentrate on the stack layout, exception handling and related structures in MSVC-compiled programs. Some familiarity with assembler, registers, calling conventions etc. is assumed.

Terms:

Stack frame: A fragment of the stack segment used by a function. Usually contains function arguments, return-to-caller address, saved registers, local variables and other data specific to this function. On x86 (and most other architectures) caller and callee stack frames are contiguous.
Frame pointer: A register or other variable that points to a fixed location inside the stack frame. Usually all data inside the stack frame is addressed relative to the frame pointer. On x86 it's usually ebp and it usually points just below the return address.
Object: An instance of a (C++) class.
Unwindable Object: A local object with auto storage-class specifier that is allocated on the stack and needs to be destructed when it goes out of scope.
Stack UInwinding: Automatic destruction of such objects that happens when the control leaves the scope due to an exception.

There are two types of exceptions that can be used in a C or C++ program.

SEH exceptions (from "Structured Exception Handling"). Also known as Win32 or system exceptions. These are exhaustively covered in the famous Matt Pietrek article[1]. They are the only exceptions available to C programs. The compiler-level support includes keywords __try, __except, __finally and a few others.
C++ exceptions (sometimes referred to as "EH"). Implemented on top of SEH, C++ exceptions allow throwing and catching of arbitrary types. A very important feature of C++ is automatic stack unwinding during exception processing, and MSVC uses a pretty complex underlying framework to ensure that it works properly in all cases.

In the following diagrams memory addresses increase from top to bottom, so the stack grows "up". It's the way the stack is represented in IDA and opposite to the most other publications.

Basic Frame Layout

The most basic stack frame looks like following:

...

Local variables

Other saved registers

Saved ebp

Return address

Function arguments

...

Note: If frame pointer omission is enabled, saved ebp might be absent.

SEH

In cases where the compiler-level SEH (__try/__except/__finally) is used, the stack layout gets a little more complicated.

SEH3 Stack Layout

When there are no __except blocks in a function (only __finally), Saved ESP is not used. Scopetable is an array of records which describe each __try block and relationships between them:

struct _SCOPETABLE_ENTRY {

DWORD EnclosingLevel;

void* FilterFunc;

void* HandlerFunc;

}

For more details on SEH implementation see[1]. To recover try blocks watch how the try level variable is updated. It's assigned a unique number per try block, and nesting is described by relationship between scopetable entries. E.g. if scopetable entry i has EnclosingLevel=j, then try block j encloses try block i. The function body is considered to have try level -1. See Appendix 1 for an example.

Buffer Overrun Protection

The Whidbey (MSVC 2005) compiler adds some buffer overrun protection for the SEH frames. The full stack frame layout in it looks like following:

SEH4 Stack Layout

The GS cookie is present only if the function was compiled with /GS switch. The EH cookie is always present. The SEH4 scopetable is basically the same as SEH3 one, only with added header:

struct _EH4_SCOPETABLE {

DWORD GSCookieOffset;

DWORD GSCookieXOROffset;

DWORD EHCookieOffset;

DWORD EHCookieXOROffset;

_EH4_SCOPETABLE_RECORD ScopeRecord[1];

};

struct _EH4_SCOPETABLE_RECORD {

DWORD EnclosingLevel;

long (*FilterFunc)();

union {

void (*HandlerAddress)();

void (*FinallyFunc)();

};

GSCookieOffset = -2 means that GS cookie is not used. EH cookie is always present. Offsets are ebp relative. Check is done the following way: (ebp+CookieXOROffset) ^ [ebp+CookieOffset] == _security_cookie Pointer to the scopetable in the stack is XORed with the _security_cookie too. Also, in SEH4 the outermost scope level is -2, not -1 as in SEH3.

C++ Exception Model Implementation

When C++ exceptions handling (try/catch) or unwindable objects are present in the function, things get pretty complex.

C++ EH Stack Layout

EH handler is different for each function (unlike the SEH case) and usually looks like this:

(VC7+)

mov eax, OFFSET __ehfuncinfo

jmp ___CxxFrameHandler

__ehfuncinfo is a structure of type FuncInfo which fully describes all try/catch blocks and unwindable objects in the function.

struct FuncInfo {

// compiler version.

// 0x19930520: up to VC6, 0x19930521: VC7.x(2002-2003), 0x19930522: VC8 (2005)

DWORD magicNumber;

// number of entries in unwind table

int maxState;

// table of unwind destructors

UnwindMapEntry* pUnwindMap;

// number of try blocks in the function

DWORD nTryBlocks;

// mapping of catch blocks to try blocks

TryBlockMapEntry* pTryBlockMap;

// not used on x86

DWORD nIPMapEntries;

// not used on x86

void* pIPtoStateMap;

// VC7+ only, expected exceptions list (function "throw" specifier)

ESTypeList* pESTypeList;

// VC8+ only, bit 0 set if function was compiled with /EHs

int EHFlags;

};

Unwind map is similar to the SEH scopetable, only without filter functions:

struct UnwindMapEntry {

int toState; // target state

void (*action)(); // action to perform (unwind funclet address)

};

Try block descriptor. Describes a try{} block with associated catches.

struct TryBlockMapEntry {

int tryLow;

int tryHigh; // this try {} covers states ranging from tryLow to tryHigh

int catchHigh; // highest state inside catch handlers of this try

int nCatches; // number of catch handlers

HandlerType* pHandlerArray; //catch handlers table

};

Catch block descriptor. Describes a single catch() of a try block.

struct HandlerType {

// 0x01: const, 0x02: volatile, 0x08: reference

DWORD adjectives;

// RTTI descriptor of the exception type. 0=any (ellipsis)

TypeDescriptor* pType;

// ebp-based offset of the exception object in the function stack.

// 0 = no object (catch by type)

int dispCatchObj;

// address of the catch handler code.

// returns address where to continues execution (i.e. code after the try block)

void* addressOfHandler;

};

List of expected exceptions (implemented but not enabled in MSVC by default, use /d1ESrt to enable).

struct ESTypeList {

// number of entries in the list

int nCount;

// list of exceptions; it seems only pType field in HandlerType is used

HandlerType* pTypeArray;

};

RTTI type descriptor. Describes a single C++ type. Used here to match the thrown exception type with catch type.

struct TypeDescriptor {

// vtable of type_info class

const void * pVFTable;

// used to keep the demangled name returned by type_info::name()

void* spare;

// mangled type name, e.g. ".H" = "int", ".?AUA@@" = "struct A", ".?AVA@@" = "class A"

char name[0];

};

Unlike SEH, each try block doesn't have a single associated state value. The compiler changes the state value not only on entering/leaving a try block, but also for each constructed/destroyed object. That way it's possible to know which objects need unwinding when an exception happens. You can still recover try blocks boundaries by inspecting the associated state range and the addresses returned by catch handlers (see Appendix 2).

Throwing C++ Exceptions

throw statements are converted into calls of _CxxThrowException(), which actually raises a Win32 (SEH) exception with the code 0xE06D7363 ('msc'|0xE0000000). The custom parameters of the Win32 exception include pointers to the exception object and its ThrowInfo structure, using which the exception handler can match the thrown exception type against the types expected by catch handlers.

struct ThrowInfo {

// 0x01: const, 0x02: volatile

DWORD attributes;

// exception destructor

void (*pmfnUnwind)();

// forward compatibility handler

int (*pForwardCompat)();

// list of types that can catch this exception.

// i.e. the actual type and all its ancestors.

CatchableTypeArray* pCatchableTypeArray;

};

struct CatchableTypeArray {

// number of entries in the following array

int nCatchableTypes;

CatchableType* arrayOfCatchableTypes[0];

};

Describes a type that can catch this exception.

struct CatchableType {

// 0x01: simple type (can be copied by memmove), 0x02: can be caught by reference only, 0x04: has virtual bases

DWORD properties;

// see above

TypeDescriptor* pType;

// how to cast the thrown object to this type

PMD thisDisplacement;

// object size

int sizeOrOffset;

// copy constructor address

void (*copyFunction)();

};

// Pointer-to-member descriptor.

struct PMD {

// member offset

int mdisp;

// offset of the vbtable (-1 if not a virtual base)

int pdisp;

// offset to the displacement value inside the vbtable

int vdisp;

};

We'll delve more into this in the next article.

Prologs and Epilogs

Instead of emitting the code for setting up the stack frame in the function body, the compiler might choose to call specific prolog and epilog functions instead. There are several variants, each used for specific function type:

Name	Type	EH Cookie	GS Cookie	Catch Handlers
_SEH_prolog/_SEH_epilog	SEH3	-	-
_SEH_prolog4/_SEH_epilog4 S	EH4	+	-
_SEH_prolog4_GS/_SEH_epilog4_GS	SEH4	+	+
_EH_prolog	C++ EH	-	-	+/-
_EH_prolog3/_EH_epilog3	C++ EH	+	-	-
_EH_prolog3_catch/_EH_epilog3	C++ EH	+	-	+
_EH_prolog3_GS/_EH_epilog3_GS	C++ EH	+	+	-
_EH_prolog3_catch_GS/_EH_epilog3_catch_GS	C++ EH	+	+	+

SEH2

Apparently was used by MSVC 1.XX (exported by crtdll.dll). Encountered in some old NT programs.

...

Saved edi

Saved esi

Saved ebx

Next SEH frame

Current SEH handler (__except_handler2)

Pointer to the scopetable

Try level

Saved ebp (of this function)

Exception pointers

Local variables

Saved ESP

Local variables

Callee EBP

Return address

Function arguments

...

Appendix I: Sample SEH Program

Let's consider the following sample disassembly.

func1 proc near

_excCode = dword ptr -28h

buf = byte ptr -24h

_saved_esp = dword ptr -18h

_exception_info = dword ptr -14h

_next = dword ptr -10h

_handler = dword ptr -0Ch

_scopetable = dword ptr -8

_trylevel = dword ptr -4

str = dword ptr 8

push ebp

mov ebp, esp

push -1

push offset _func1_scopetable

push offset _except_handler3

mov eax, large fs:0

push eax

mov large fs:0, esp

add esp, -18h

push ebx

push esi

push edi

; --- end of prolog ---

mov [ebp+_trylevel], 0 ;trylevel -1 -> 0: beginning of try block 0

mov [ebp+_trylevel], 1 ;trylevel 0 -> 1: beginning of try block 1

mov large dword ptr ds:123, 456

mov [ebp+_trylevel], 0 ;trylevel 1 -> 0: end of try block 1

jmp short _endoftry1

_func1_filter1: ; __except() filter of try block 1

mov ecx, [ebp+_exception_info]

mov edx, [ecx+EXCEPTION_POINTERS.ExceptionRecord]

mov eax, [edx+EXCEPTION_RECORD.ExceptionCode]

mov [ebp+_excCode], eax

mov ecx, [ebp+_excCode]

xor eax, eax

cmp ecx, EXCEPTION_ACCESS_VIOLATION

setz al

retn

_func1_handler1: ; beginning of handler for try block 1

mov esp, [ebp+_saved_esp]

push offset aAccessViolatio ; "Access violation"

call _printf

add esp, 4

mov [ebp+_trylevel], 0 ;trylevel 1 -> 0: end of try block 1

_endoftry1:

mov edx, [ebp+str]

push edx

lea eax, [ebp+buf]

push eax

call _strcpy

add esp, 8

mov [ebp+_trylevel], -1 ; trylevel 0 -> -1: end of try block 0

call _func1_handler0 ; execute __finally of try block 0

jmp short _endoftry0

_func1_handler0: ; __finally handler of try block 0

push offset aInFinally ; "in finally"

call _puts

add esp, 4

retn

_endoftry0:

; --- epilog ---

mov ecx, [ebp+_next]

mov large fs:0, ecx

pop edi

pop esi

pop ebx

mov esp, ebp

pop ebp

retn

func1 endp

_func1_scopetable

;try block 0

dd -1 ;EnclosingLevel

dd 0 ;FilterFunc

dd offset _func1_handler0 ;HandlerFunc

;try block 1

dd 0 ;EnclosingLevel

dd offset _func1_filter1 ;FilterFunc

dd offset _func1_handler1 ;HandlerFunc

The try block 0 has no filter, therefore its handler is a __finally{} block. EnclosingLevel of try block 1 is 0, so it's placed inside try block 0. Considering this, we can try to reconstruct the function structure:

void func1 (char* str)

{

char buf[12];

__try // try block 0

{

__try // try block 1

{

*(int*)123=456;

}

__except(GetExceptCode() == EXCEPTION_ACCESS_VIOLATION)

{

printf("Access violation");

}

strcpy(buf,str);

}

__finally

{

puts("in finally");

}

Appendix II: Sample Program with C++ Exceptions

func1 proc near

_a1 = dword ptr -24h

_exc = dword ptr -20h

e = dword ptr -1Ch

a2 = dword ptr -18h

a1 = dword ptr -14h

_saved_esp = dword ptr -10h

_next = dword ptr -0Ch

_handler = dword ptr -8

_state = dword ptr -4

push ebp

mov ebp, esp

push 0FFFFFFFFh

push offset func1_ehhandler

mov eax, large fs:0

push eax

mov large fs:0, esp

push ecx

sub esp, 14h

push ebx

push esi

push edi

mov [ebp+_saved_esp], esp

; --- end of prolog ---

lea ecx, [ebp+a1]

call A::A(void)

mov [ebp+_state], 0 ; state -1 -> 0: a1 constructed

mov [ebp+a1], 1 ; a1.m1 = 1

mov byte ptr [ebp+_state], 1 ; state 0 -> 1: try {

lea ecx, [ebp+a2]

call A::A(void)

mov [ebp+_a1], eax

mov byte ptr [ebp+_state], 2 ; state 2: a2 constructed

mov [ebp+a2], 2 ; a2.m1 = 2

mov eax, [ebp+a1]

cmp eax, [ebp+a2] ; a1.m1 == a2.m1?

jnz short loc_40109F

mov [ebp+_exc], offset aAbc ; _exc = "abc"

push offset __TI1?PAD ; char *

lea ecx, [ebp+_exc]

push ecx

call _CxxThrowException ; throw "abc";

loc_40109F:

mov byte ptr [ebp+_state], 1 ; state 2 -> 1: destruct a2

lea ecx, [ebp+a2]

call A::~A(void)

jmp short func1_try0end

; catch (char * e)

func1_try0handler_pchar:

mov edx, [ebp+e]

push edx

push offset aCaughtS ; "Caught %s\n"

call ds:printf ;

add esp, 8

mov eax, offset func1_try0end

retn

; catch (...)

func1_try0handler_ellipsis:

push offset aCaught___ ; "Caught ...\n"

call ds:printf

add esp, 4

mov eax, offset func1_try0end

retn

func1_try0end:

mov [ebp+_state], 0 ; state 1 -> 0: }//try

push offset aAfterTry ; "after try\n"

call ds:printf

add esp, 4

mov [ebp+_state], -1 ; state 0 -> -1: destruct a1

lea ecx, [ebp+a1]

call A::~A(void)

; --- epilog ---

mov ecx, [ebp+_next]

mov large fs:0, ecx

pop edi

pop esi

pop ebx

mov esp, ebp

pop ebp

retn

func1 endp

func1_ehhandler proc near

mov eax, offset func1_funcinfo

jmp __CxxFrameHandler

func1_ehhandler endp

func1_funcinfo

dd 19930520h ; magicNumber

dd 4 ; maxState

dd offset func1_unwindmap ; pUnwindMap

dd 1 ; nTryBlocks

dd offset func1_trymap ; pTryBlockMap

dd 0 ; nIPMapEntries

dd 0 ; pIPtoStateMap

dd 0 ; pESTypeList

func1_unwindmap

dd -1

dd offset func1_unwind_1tobase ; action

dd 0 ; toState

dd 0 ; action

dd 1 ; toState

dd offset func1_unwind_2to1 ; action

dd 0 ; toState

dd 0 ; action

func1_trymap

dd 1 ; tryLow

dd 2 ; tryHigh

dd 3 ; catchHigh

dd 2 ; nCatches

dd offset func1_tryhandlers_0 ; pHandlerArray

dd 0

func1_tryhandlers_0

dd 0 ; adjectives

dd offset char * `RTTI Type Descriptor' ; pType

dd -1Ch ; dispCatchObj

dd offset func1_try0handler_pchar ; addressOfHandler

dd 0 ; adjectives

dd 0 ; pType

dd 0 ; dispCatchObj

dd offset func1_try0handler_ellipsis ; addressOfHandler

func1_unwind_1tobase proc near

a1 = byte ptr -14h

lea ecx, [ebp+a1]

call A::~A(void)

retn

func1_unwind_1tobase endp

func1_unwind_2to1 proc near

a2 = byte ptr -18h

lea ecx, [ebp+a2]

call A::~A(void)

retn

func1_unwind_2to1 endp

Let's see what we can find out here. The maxState field in FuncInfo structure is 4 which means we have four entries in the unwind map, from 0 to 3. Examining the map, we see that the following actions are executed during unwinding:

state 3 -> state 0 (no action)
state 2 -> state 1 (destruct a2)
state 1 -> state 0 (no action)
state 0 -> state -1 (destruct a1)

Checking the try map, we can infer that states 1 and 2 correspond to the try block body and state 3 to the catch blocks bodies. Thus, change from state 0 to state 1 denotes the beginning of try block, and change from 1 to 0 its end. From the function code we can also see that -1 -> 0 is construction of a1, and 1 -> 2 is construction of a2. So the state diagram looks like this:

Where did the arrow 1->3 come from? We cannot see it in the function code or FuncInfo structure since it's done by the exception handler. If an exception happens inside try block, the exception handler first unwinds the stack to the tryLow value (1 in our case) and then sets state value to tryHigh+1 (2+1=3) before calling the catch handler.

The try block has two catch handlers. The first one has a catch type (char*) and gets the exception object on the stack (-1Ch = e). The second one has no type (i.e. ellipsis catch). Both handlers return the address where to resume execution, i.e. the position just after the try block. Now we can recover the function code:

void func1 ()

{

A a1;

a1.m1 = 1;

try {

A a2;

a2.m1 = 2;

if (a1.m1 == a1.m2) throw "abc";

}

catch(char* e)

{

printf("Caught %s\n",e);

}

catch(...)

{

printf("Caught ...\n");

}

printf("after try\n");

}

Appendix III: IDC Helper Scripts

I wrote an IDC script to help with the reversing of MSVC programs. It scans the whole program for typical SEH/EH code sequences and comments all related structures and fields. Commented are stack variables, exception handlers, exception types and other. It also tries to fix function boundaries that are sometimes incorrectly determined by IDA. You can download it from MS SEH/EH Helper.

Links and References

[1] Matt Pietrek. A Crash Course on the Depths of Win32 Structured Exception Handling.
http://www.microsoft.com/msj/0197/exception/exception.aspx
Still THE definitive guide on the implementation of SEH in Win32.

[2] Brandon Bray. Security Improvements to the Whidbey Compiler.
http://blogs.msdn.com/branbray/archive/2003/11/11/51012.aspx
Short description on changes in the stack layout for cookie checks.

[3] Chris Brumme. The Exception Model.
http://blogs.msdn.com/cbrumme/archive/2003/10/01/51524.aspx
Mostly about .NET exceptions, but still contains a good deal of information about SEH and C++ exceptions.

[4] Vishal Kochhar. How a C++ compiler implements exception handling.
http://www.codeproject.com/cpp/exceptionhandler.asp
An overview of C++ exceptions implementation.

[5] Calling Standard for Alpha Systems. Chapter 5. Event Processing.
http://www.cs.arizona.edu/computer.help/policy/DIGITAL_unix/AA-PY8AC-TET1_html/callCH5.html
Win32 takes a lot from the way Alpha handles exceptions and this manual has a very detailed description on how it happens.

Structure definitions and flag values were also recovered from the following sources:

VC8 CRT debug information (many structure definitions)
VC8 assembly output (/FAs)
VC8 WinCE CRT source

Reversing Microsoft Visual C++ Part II: Classes, Methods and RTTI

Abstract

Microsoft Visual C++ is the most widely used compiler for Win32 so it is important for the Win32 reverser to be familiar with its inner working. Being able to recognize the compiler-generated glue code helps to quickly concentrate on the actual code written by the programmer. It also helps in recovering the high-level structure of the program.

In part II of this 2-part article (see also: Part I: Exception Handling), I will cover how C++ machinery is implemented in MSVC, including classes layout, virtual functions, RTTI. Familiarity with basic C++ and assembly language is assumed.

Basic Class Layout

To illustrate the following material, let's consider this simple example:

class A

{

int a1;

public:

virtual int A_virt1();

virtual int A_virt2();

static void A_static1();

void A_simple1();

};

class B

{

int b1;

int b2;

public:

virtual int B_virt1();

virtual int B_virt2();

};

class C: public A, public B

{

int c1;

public:

virtual int A_virt2();

virtual int B_virt2();

};

In most cases MSVC lays out classes in the following order:

1. Pointer to virtual functions table (_vtable_ or _vftable_), added only when the class has virtual methods and no suitable table from a base class can be reused.
2. Base classes
3. Class members

Virtual function tables consist of addresses of virtual methods in the order of their first appearance. Addresses of overloaded functions replace addresses of functions from base classes.

Thus, the layouts for our three classes will look like following:

class A size(8):

+---

0 | {vfptr}

4 | a1

+---

A's vftable:

0 | &A::A_virt1

4 | &A::A_virt2

class B size(12):

+---

0 | {vfptr}

4 | b1

8 | b2

+---

B's vftable:

0 | &B::B_virt1

4 | &B::B_virt2

class C size(24):

+---

| +--- (base class A)

0 | | {vfptr}

4 | | a1

| +---

| +--- (base class B)

8 | | {vfptr}

12 | | b1

16 | | b2

| +---

20 | c1

+---

C's vftable for A:

0 | &A::A_virt1

4 | &C::A_virt2

C's vftable for B:

0 | &B::B_virt1

4 | &C::B_virt2

The above diagram was produced by the VC8 compiler using an undocumented switch. To see the class layouts produced by the compiler, use: -d1reportSingleClassLayout to see the layout of a single class -d1reportAllClassLayout to see the layouts of all classes (including internal CRT classes) The layouts are dumped to stdout.

As you can see, C has two vftables, since it has inherited two classes which both already had virtual functions. Address of C::A_virt2 replaces address of A::A_virt2 in C's vftable for A, and C::B_virt2 replaces B::B_virt2 in the other table.

Calling Conventions and Class Methods

All class methods in MSVC by default use _thiscall_ convention. Class instance address (_this_ pointer) is passed as a hidden parameter in the ecx register. In the method body the compiler usually tucks it away immediately in some other register (e.g. esi or edi) and/or stack variable. All further adressing of the class members is done through that register and/or variable. However, when implementing COM classes, _stdcall_ convention is used. The following is an overview of the various class method types.

1) Static Methods
Static methods do not need a class instance, so they work the same way as common functions. No _this_ pointer is passed to them. Thus it's not possible to reliably distinguish static methods from simple functions. Example:

A::A_static1();

call A::A_static1

2) Simple Methods
Simple methods need a class instance, so _this_ pointer is passed to them as a hidden first parameter, usually using _thiscall_ convention, i.e. in _ecx_ register. When the base object is not situated at the beginning of the derived class, _this_ pointer needs to be adjusted to point to the actual beginning of the base subobject before calling the function. Example:

;pC->A_simple1(1);

;esi = pC

push 1

mov ecx, esi

call A::A_simple1

;pC->B_simple1(2,3);

;esi = pC

lea edi, [esi+8] ;adjust this

push 3

push 2

mov ecx, edi

call B::B_simple1

As you see, _this_ pointer is adjusted to point to the B subobject before calling B's method.

3) Virtual Methods
To call a virtual method the compiler first needs to fetch the function address from the _vftable_ and then call the function at that address same way as a simple method (i.e. passing _this_ pointer as an implicit parameter). Example:

;pC->A_virt2()

;esi = pC

mov eax, [esi] ;fetch virtual table pointer

mov ecx, esi

call [eax+4] ;call second virtual method

;pC->B_virt1()

;edi = pC

lea edi, [esi+8] ;adjust this pointer

mov eax, [edi] ;fetch virtual table pointer

mov ecx, edi

call [eax] ;call first virtual method

4) Constructors and Destructors
Constructors and destructors work similar to a simple method: they get an implicit _this_ pointer as the first parameter (e.g. ecx in case of _thiscall_ convention). Constructor returns the _this_ pointer in eax, even though formally it has no return value.

RTTI Implementation

RTTI (Run-Time Type Identification) is special compiler-generated information which is used to support C++ operators like dynamic_cast<> and typeid(), and also for C++ exceptions. Due to its nature, RTTI is only required (and generated) for polymorphic classes, i.e. classes with virtual functions.

MSVC compiler puts a pointer to the structure called "Complete Object Locator" just before the vftable. The structure is called so because it allows compiler to find the location of the complete object from a specific vftable pointer (since a class can have several of them). COL looks like following:

struct RTTICompleteObjectLocator

{

DWORD signature; //always zero ?

DWORD offset; //offset of this vtable in the complete class

DWORD cdOffset; //constructor displacement offset

struct TypeDescriptor* pTypeDescriptor; //TypeDescriptor of the complete class

struct RTTIClassHierarchyDescriptor* pClassDescriptor; //describes inheritance hierarchy

};

Class Hierarchy Descriptor describes the inheritance hierarchy of the class. It is shared by all COLs for a class.

struct RTTIClassHierarchyDescriptor

{

DWORD signature; //always zero?

DWORD attributes; //bit 0 set = multiple inheritance, bit 1 set = virtual inheritance

DWORD numBaseClasses; //number of classes in pBaseClassArray

struct RTTIBaseClassArray* pBaseClassArray;

};

Base Class Array describes all base classes together with information which allows compiler to cast the derived class to any of them during execution of the _dynamic_cast_ operator. Each entry (Base Class Descriptor) has the following structure:

struct RTTIBaseClassDescriptor

{

struct TypeDescriptor* pTypeDescriptor; //type descriptor of the class

DWORD numContainedBases; //number of nested classes following in the Base Class Array

struct PMD where; //pointer-to-member displacement info

DWORD attributes; //flags, usually 0

};

struct PMD

{

int mdisp; //member displacement

int pdisp; //vbtable displacement

int vdisp; //displacement inside vbtable

};

The PMD structure describes how a base class is placed inside the complete class. In the case of simple inheritance it is situated at a fixed offset from the start of object, and that value is the _mdisp_ field. If it's a virtual base, an additional offset needs to be fetched from the vbtable. Pseudo-code for adjusting _this_ pointer from derived class to a base class looks like the following:

//char* pThis; struct PMD pmd;

pThis+=pmd.mdisp;

if (pmd.pdisp!=-1)

{

char *vbtable = pThis+pmd.pdisp;

pThis += *(int*)(vbtable+pmd.vdisp);

}

For example, the RTTI hierarchy for our three classes looks like this:

RTTI hierarchy for our example classes

Extracting Information

1) RTTI
If present, RTTI is a valuable source of information for reversing. From RTTI it's possible to recover class names, inheritance hierarchy, and in some cases parts of the class layout. My RTTI scanner script shows most of that information. (see Appendix I)

2) Static and Global Initializers
Global and static objects need to be initialized before the main program starts. MSVC implements that by generating initializer funclets and putting their addresses in a table, which is processed during CRT startup by the _cinit function. The table usually resides in the beginning of .data section. A typical initializer looks like following:

_init_gA1:

mov ecx, offset _gA1

call A::A()

push offset _term_gA1

call _atexit

pop ecx

retn

_term_gA1:

mov ecx, offset _gA1

call A::~A()

retn

Thus, from this table way we can find out:

Global/static objects addresses
Their constructors
Their destructors

See also MSVC _#pragma_ directive _init_seg_ [5].

3) Unwind Funclets
If any automatic objects are created in a function, VC++ compiler automatically generates exception handling structures which ensure deletion of those objects in case an exception happens. See Part I for a detailed description of C++ exception implementation. A typical unwind funclet destructs an object on the stack:

unwind_1tobase: ; state 1 -> -1

lea ecx, [ebp+a1]

jmp A::~A()

By finding the opposite state change inside the function body or just the first access to the same stack variable, we can also find the constructor:

lea ecx, [ebp+a1]

call A::A()

mov [ebp+__$EHRec$.state], 1

For the objects constructed using new() operator, the unwind funclet ensures deletion of allocated memory in case the constructor fails:

unwind_0tobase: ; state 0 -> -1

mov eax, [ebp+pA1]

push eax

call operator delete(void *)

pop ecx

retn

In the function body:

;A* pA1 = new A();

push

call operator new(uint)

add esp, 4

mov [ebp+pA1], eax

test eax, eax

mov [ebp+__$EHRec$.state], 0; state 0: memory allocated but object is not yet constructed

jz short @@new_failed

mov ecx, eax

call A::A()

mov esi, eax

jmp short @@constructed_ok

@@new_failed:

xor esi, esi

@@constructed_ok:

mov [esp+14h+__$EHRec$.state], -1

;state -1: either object was constructed successfully or memory allocation failed

;in both cases further memory management is done by the programmer

Another type of unwind funclets is used in constructors and destructors. It ensures destruction of the class members in case of exception. In this case the funclets use the _this_ pointer, which is kept in a stack variable:

unwind_2to1:

mov ecx, [ebp+_this] ; state 2 -> 1

add ecx, 4Ch

jmp B1::~B1

Here the funclet destructs a class member of type B1 at the offset 4Ch. Thus, from unwind funclets we can find out:

Stack variables representing C++ objects or pointers to objects allocated with _operator new_.
Their destructors
Their constructors
in case of new'ed objects, their size

4) Constructors / Destructors Recursion
This rule is simple: constructors call other constructors (of base classes and member variables) and destructors call other destructors. A typical constructor does the following:

Call constructors of the base classes.
Call constructors of complex class members.
Initialize vfptr(s) if the class has virtual functions
Execute the constructor body written by the programmer.

Typical destructor works almost in the reverse order:

Initialize vfptr if the class has virtual functions
Execute the destructor body written by the programmer.
Call destructors of complex class members
Call destructors of base classes

Another distinctive feature of destructors generated by MSVC is that their _state_ variable is usually initialized with the highest value and then gets decremented with each destructed subobject, which make their identification easier. Be aware that simple constructors/destructors are often inlined by MSVC. That's why you can often see the vftable pointer repeatedly reloaded with different pointers in the same function.

5) Array Construction Destruction
The MSVC compiler uses a helper function to construct and destroy an array of objects. Consider the following code:

A* pA = new A[n];

delete [] pA;

It is translated into the following pseudocode:

array = new char(sizeof(A)*n+sizeof(int))

if (array)

{

*(int*)array=n; //store array size in the beginning

'eh vector constructor iterator'(array+sizeof(int),sizeof(A),count,&A::A,&A::~A);

}

pA = array;

'eh vector destructor iterator'(pA,sizeof(A),count,&A::~A);

If A has a vftable, a 'vector deleting destructor' is invoked instead when deleting the array:

;pA->'vector deleting destructor'(3);

mov ecx, pA

push 3 ; flags: 0x2=deleting an array, 0x1=free the memory

call A::'vector deleting destructor'

If A's destructor is virtual, it's invoked virtually:

mov ecx, pA

push 3

mov eax, [ecx] ;fetch vtable pointer

call [eax] ;call deleting destructor

Consequently, from the vector constructor/destructor iterator calls we can determine:

addresses of arrays of objects
their constructors
their destructors
class sizes

6) Deleting Destructors
When class has a virtual destructor, compiler generates a helper function - deleting destructor. Its purpose is to make sure that a proper _operator delete_ gets called when destructing a class. Pseudo-code for a deleting destructor looks like following:

virtual void * A::'scalar deleting destructor'(uint flags)

{

this->~A();

if (flags&1) A::operator delete(this);

};

The address of this function is placed into the vftable instead of the destructor's address. This way, if another class overrides the virtual destructor, _operator delete_ of that class will be called. Though in real code _operator delete_ gets overriden quite rarely, so usually you see a call to the default delete(). Sometimes compiler can also generate a vector deleting destructor. Its code looks like this:

virtual void * A::'vector deleting destructor'(uint flags)

{

if (flags&2) //destructing a vector

{

array = ((int*)this)-1; //array size is stored just before the this pointer

count = array[0];

'eh vector destructor iterator'(this,sizeof(A),count,A::~A);

if (flags&1) A::operator delete(array);

}

else {

this->~A();

if (flags&1) A::operator delete(this);

}

};

I skipped most of the details on implementation of classes with virtual bases since they complicate things quite a bit and are rather rare in the real world. Please refer to the article by Jan Gray[1]. It's very detailed, if a bit heavy on Hungarian notation. The article [2] describes an example of the virtual inheritance implementation in MSVC. See also some of the MS patents [3] for more details.

Appendix I: ms_rtti4.idc

This is a script I wrote for parsing RTTI and vftables. You can download the scripts associated with both this article and the previous article from Microsoft VC++ Reversing Helpers. The script features:

Parses RTTI structures and renames vftables to use the corresponding class names.
For some simple cases, identifies and renames constructors and destructors.
Outputs a file with the list of all vftables with referencing functions and class hierarchy.

Usage: after the initial analysis finishes, load ms_rtti4.idc. It will ask if you want to scan the exe for the vtables. Be aware that it can be a lengthy process. Even if you skip the scanning, you can still parse vtables manually. If you do choose to scan, the script will try to identify all vtables with RTII, rename them, and identify and rename constructors and destructors. In some cases it will fail, especially with virtual inheritance. After scanning, it will open the text file with results.

After the script is loaded, you can use the following hotkeys to parse some of the MSVC structures manually:
Alt-F8 - parse a vtable. The cursor should be at the beginning of the vtable. If there is RTTI, the script will use the class name from it. If there is none, you can enter the class name manually and the script will rename the vtable. If there is a virtual destructor which it can identify, the script will rename it too.
Alt-F7 - parse FuncInfo. FuncInfo is the structure present in functions which have objects allocated on the stack or use exception handling. Its address is passed to _CxxFrameHandler in the function's exception handler:

· mov eax, offset FuncInfo1

· jmp _CxxFrameHandler

In most cases it is identified and parsed automatically by IDA, but my script provides more information. You can also use ms_ehseh.idc from the first part of this article to parse all FuncInfos in the file.
Use the hotkey with cursor placed on the start of the FuncInfo structure.

Alt-F9 - parse throw info. Throw info is a helper structure used by _CxxThrowException to implement the _throw_ operator. Its address is the second argument to _CxxThrowException:

· lea ecx, [ebp+e]

· call E::E()

· push offset ThrowInfo_E

· lea eax, [ebp+e]

· push eax

· call _CxxThrowException

Use the hotkey with the cursor placed on the start of the throw info structure. The script will parse the structure and add a repeatable comment with the name of the thrown class. It will also identify and rename the exception's destructor and copy constructor.

Appendix II: Practical Recovery of a Class Structure

Our subject will be MSN Messenger 7.5 (msnmsgr.exe version 7.5.324.0, size 7094272). It makes heavy use of C++ and has plenty of RTTI for our purposes. Let's consider two vftables, at .0040EFD8 and .0040EFE0. The complete RTTI structures hierarchy for them looks like following:

RTTI hierarchy for MSN Messenger 7.5

So, these two vftables both belong to one class - CContentMenuItem. By checking its Base Class Descriptors we can see that:

CContentMenuItem contains three bases that follow it in the array - i.e. CDownloader, CNativeEventSink and CNativeEventSource.
CDownloader contains one base - CNativeEventSink.
Hence, CContentMenuItem inherits directly from CDownloader and CNativeEventSource, and CDownloader in turn inherits from CNativeEventSink.
CDownloader is situated in the beginning of the complete object, and CNativeEventSource is at the offset 0x24.

So we can conclude that the first vftable lists methods of CNativeEventSource and the second one of either CDownloader or CNativeEventSink (if neither of them had virtual methods, CContentMenuItem would reuse the vftable of CNativeEventSource). Now let's check what refers to these tables. They both are referred by two functions, at .052B5E0 and .052B547. (That reinforces the fact that they both belong to one class.) Moreover, if we look at the beginning of the function at .052B547, we see the _state_ variable initialized with 6, which means that that function is the destructor. As a class can have only one destructor, we can conclude that .052B5E0 is its constructor. Let's looks closer at it:

CContentMenuItem::CContentMenuItem proc near

this = esi

push this

push edi

mov this, ecx

call sub_4CA77A

lea edi, [this+24h]

mov ecx, edi

call sub_4CBFDB

or dword ptr [this+48h], 0FFFFFFFFh

lea ecx, [this+4Ch]

mov dword ptr [this], offset const CContentMenuItem::'vftable'{for 'CContentMenuItem'}

mov dword ptr [edi], offset const CContentMenuItem::'vftable'{for 'CNativeEventSource'}

call sub_4D8000

lea ecx, [this+50h]

call sub_4D8000

lea ecx, [this+54h]

call sub_4D8000

lea ecx, [this+58h]

call sub_4D8000

lea ecx, [this+5Ch]

call sub_4D8000

xor eax, eax

mov [this+64h], eax

mov [this+68h], eax

mov [this+6Ch], eax

pop edi

mov dword ptr [this+60h], offset const CEventSinkList::'vftable'

mov eax, this

pop this

retn

sub_52B5E0 endp

The first thing compiler does after prolog is copying _this_ pointer from ecx to esi, so all further addressing is done based on esi. Before initializing vfptrs it calls two other functions; those must be constructors of the base classes - in our case CDownloader and CNativeEventSource. We can confirm that by going inside each of the functions - first one initializes its vfptr field with CDownloader::'vftable' and the second with CNativeEventSource::'vftable'. We can also investigate CDownloader's constructor further - it calls constructor of its base class, CNativeEventSink.

Also, the _this_ pointer passed to the second function is taken from edi, which points to this+24h. According to our class structure diagram it's the location of the CNativeEventSource subobject. This is another confirmation that the second function being called is the constructor of CNativeEventSource.

After calling base constructors, the vfptrs of the base objects are overwritten with CContentMenuItem's implementations - which means that CContentMenuItem overrides some of the virtual methods of the base classes (or adds its own). (If needed, we can compare the tables and check which pointers have been changed or added - those will be new implementations by CContentMenuItem.)

Next we see several function calls to .04D8000 with _ecx_ set to this+4Ch to this+5Ch - apparently some member variables are initialized. How can we know whether that function is a compiler-generated constructor call or an initializer function written by the programmer? There are several hints that it's a constructor.

The function uses _thiscall_ convention and it is the first time these fields are accessed.
The fields are initialized in the order of increasing addresses.

To be sure we can also check the unwind funclets in the destructor - there we can see the compiler-generated destructor calls for these member variables.

This new class doesn't have virtual methods and thus no RTTI, so we don't know its real name. Let's name it RefCountedPtr. As we have already determined, 4D8000 is its constructor. The destructor we can find out from the CContentMenuItem destructor's unwind funclets - it's at 63CCB4.

Going back to the CContentMenuItem constructor, we see three fields initialized with 0 and one with a vftable pointer. This looks like an inlined constructor for a member variable (not a base class, since a base class would be present in the inheritance tree). From the used vftable's RTTI we can see that it's an instance of CEventSinkList template.

Now we can write a possible declaration for our class.

class CContentMenuItem: public CDownloader, public CNativeEventSource

{

/* 00 CDownloader */

/* 24 CNativeEventSource */

/* 48 */ DWORD m_unknown48;

/* 4C */ RefCountedPtr m_ptr4C;

/* 50 */ RefCountedPtr m_ptr50;

/* 54 */ RefCountedPtr m_ptr54;

/* 58 */ RefCountedPtr m_ptr58;

/* 5C */ RefCountedPtr m_ptr5C;

/* 60 */ CEventSinkList m_EventSinkList;

/* size = 70? */

};

We can't know for sure that the field at offset 48 is not a part of CNativeEventSource; but since it wasn't accessed in CNativeEventSource constructor, it is most probably a part of CContentMenuItem. The constructor listing with renamed methods and class structure applied:

public: __thiscall CContentMenuItem::CContentMenuItem(void) proc near

push this

push edi

mov this, ecx

call CDownloader::CDownloader(void)

lea edi, [this+CContentMenuItem._CNativeEventSource]

mov ecx, edi

call CNativeEventSource::CNativeEventSource(void)

or [this+CContentMenuItem.m_unknown48], -1

lea ecx, [this+CContentMenuItem.m_ptr4C]

mov [this+CContentMenuItem._CDownloader._vfptr], offset const CContentMenuItem::'vftable'{for 'CContentMenuItem'}

mov [edi+CNativeEventSource._vfptr], offset const CContentMenuItem::'vftable'{for 'CNativeEventSource'}