In this essay I am going to show you how the most commonly used C++ compilers (MSVC and Borland) use the stack. Beginners will learn how the stack is used, what the function stack frame is, what the stack frame pointer is and how to use this information in order to get a function stack trace. I hope that the more advanced readers will find some interesting information, too. As an example, a simple class named
StackDumper is described.
StackDumper has a method that browses the thread's stack and allows you to save in a text file the names (or addresses in the worst case) of all functions executed before the
foo is called. As a side effect of this functionality you will be able to ask the
foo is called from
The task became a little bit complicated by my intention to create a class that is not dependent on any particular compiler. If you download the article demo you will find three projects - for MSVC++ 6, for Borland C++ 5.02 and for Borland C++ Bulder 6. If you are using the VC++ compiler there is an easy and already described way to implement this functionality. John Panzer has published in the C/C++ Users Journal, January 1999, his essay "Automatic code Instrumentation". In brief, he uses the
/Gh compiler switch to force the compiler to generate a call to a
_penter function at the start of each client function. From inside
_penter the address of the caller is retrieved and stored in a parallel stack. It also records the function entry time. Afterwards the original return address of the caller is replaced by the address of a user function. This function is used for profiling purposes. When it is called it records the function exit time and restores the original return address of the caller. This is the main idea. You can download the essay from here.
Obviously this approach uses a Microsoft-specific compiler switch. Due to the parallel stack support special attention should be paid to the case where an exception occurs in one of the profiled functions. The compiler generates a hidden call to the
_penter function in all user functions, which is not very flexible because you may not want to collect information about every function. The same functionality can be implemented by the well known concept of stack backtrace. The idea is to use the stack frame that each function builds on the stack to trace the calling sequence. This approach is compiler independent (at least it works with both MSVC and Borland compilers), but it also imposes some limitations. Most compilers provide options to compile a function without prolog and epilog code. For example functions declared with
__declspec(naked) attribute will be compiled by MSVC++ and Borland C++ compilers without prolog and epilog code. The same effect can be reached by specifying some compiler options for optimizations. For example the
/Oy switch in the Microsoft compiler suppresses the creation of frame pointers on the call stack. In this essay I take for granted that the program is compiled without any optimisations or "special" compiler switches.
1. How the stack is used?
1.1. Stack frame and frame pointer
Here is a brief explanation for those of you who are not familiar with the way the stack is used during function calls. The more advanced readers can skip this section. When a function is called its arguments are pushed onto the stack in an order that depends on the calling convention. Then the
instruction is executed. It pushes the return address onto the stack. The first instruction of the function is
- the base pointer is pushed. The stack pointer is moved to the
register, then the
register is decremented to make room on the stack for the local variables of the function. So for every called function the following information is built on the stack:
This information is called the "stack frame". The register
has a special meaning. It is called the "frame pointer". The frame pointer is initialized at the function start by the standard prolog code and stays unchanged during function execution. The value of the previous function frame pointer is restored when the current function exits (epilog). How is the frame pointer used? The compiler uses the frame pointer to refer to local variables and parameters of a function (if any).
*(ebp) is the value of the frame pointer of the caller;
*(ebp + 4) is the return address (the place in the caller body where execution will continue after the callee returns);
*(ebp + 4 + 4 * i) is the value of i-th function argument (1 based, assuming that each argument takes 4 bytes);
*(ebp - offset) refers to local variables.
You can see this operation in the following simple example:
int func(int nArg1, int nArg2)
The Borland C++ 5.02 compiler generates the following assembly code (__cdecl calling convention - the arguments are pushed on the stack from right to left, the caller cleans the stack. More information about calling convention can be found here):
Dissassembly of main:
00401110 55 push ebp <--store the ebp register
on the stack
00401111 8B EC mov ebp, esp <--current stack pointer
00401113 6A 01 push 1 <--the arguments of func
are pushed on the stack
from right to left
00401115 6A 00 push 0
00401117 E8 EC FF FF FF call func <--this call pushes the
return address (0040111C)
on the stack
0040111C 83 C4 08 add esp, 8 <-- __cdecl calling
=> caller has to clean
the space that arguments
used on the stack. Every
push decrements the stack
pointer and every pop
increments it by the size
of the operand. Two ints
are pushed on the stack
=> the stack pointer
has to be increased by 8
0040111F B8 01 00 00 00 mov eax, 1 <-- the return value of
main goes in eax
00401124 5D pop ebp <-- ebp is restored
00401125 C3 ret
Dissassembly of func:
00401108 55 push ebp |<-- prolog.
00401109 8B EC mov ebp, esp |
0040110B 83 C4 F4 add esp, -0x0c <-- make room for 3 * 4
bytes for the local
variable n on the
0040110E 8B 45 08 mov eax, [ebp+0x08] <-- move the return
value (nArg1) to eax
00401111 8B E5 mov esp, ebp |<-- epilog. epb
contains the caller
00401113 5D pop ebp |
00401114 C3 ret
It is not very hard to see how the stack frame of a function can be used to get the caller address of this function. As I have mentioned before
*(ebp + 4)
points to the return address of the function. This address is inside the body of the caller.
1.2.1. How to get the starting address of the caller?
Approach 1: Prolog searching
We suppose that all functions are compiled with a prolog and an epilog. Having an address inside the function, we just have to search for the byte sequence
55 8B EC
. As you can see from the dissassembly above these are the opcodes of the prolog. Let's call them the "prolog signature". Unfortunately there is a problem. The same sequence of bytes could appear in an instruction as an operand. For example the instruction
mov eax, EC8B55h
has the following instruction encoding:
B8 55 8B EC 00
. Obviously when we are searching the prolog signature byte by byte we will find the signature somewhere inside the mov instruction.
Note: In my experience the method of searching the prolog signature in most cases works fine. To keep things simple you can skip the next section. Note end.
Well, I don't have an elegant solution to the above problem. For that reason I am going to kill a mosquito with a nuclear bomb :-).
Approach 2: Backwards disassembling
Given an address inside the body of a function, it is easy enough to find the address of the previous instrucion. This task cannot be solved without knowing the instruction format. We need a disassembler to do this. You can find in the source a function called
#define MAX_INTEL_INSTRUCTION_LEN 15
DWORD FindAddressOfprevInstruction(DWORD EIP, PUCHAR instr)
DWORD Addr = EIP-MAX_INTEL_INSTRUCTION_LEN;
unsigned long res;
prevAddr = Addr;
res = Disasm32(&Addr, instr);
} while(Addr < EIP);
I think this function is straightforward. Having this function it is easy to find the prolog signature in a more precise way - you just have to disassemble backwards until you find the prolog signature.
1.2.2. Tracing the stack
As you can see from Figure 1. every functions stack frame contains a pointer to the callers frame which contains a pointer to its caller frame and so on. In fact we have a list of stack frames which can be used to find the callstack of every called function. But still there is an important question: When do we have to stop stack browsing, or, in other words, where does this list begin? I think we can find the answer if we take a look at the process and thread starting routines which resides in
When Windows creates a process and its main thread it performs an internal call to the
CreateProcess in turn invokes an internal routine in
BaseProcessStart. Here is the dissasembly (under the condition that you have
77e8d2e4 xor ebp,ebp <-- Look here!
77e8d2e6 push eax
77e8d2e7 push 0x0 <-- This can be interpreted as a return address
77e8d2e9 push ebp <-- This is a normal stack frame. But the
previous frame ptr is zeroed a few lines
77e8d2ea mov ebp,esp
77e8d2ec push 0xff
77e8d323 call dword ptr [ebp+0x8] <-- call the entry point of our
process(for example mainCRTStartup)
77e8d326 jmp KERNEL32!BaseProcessStart+0x3d (77eb6624)
Similar things are happening in each call to the
77e964cb xor ebp,ebp
77e964cd push ebx
77e964ce push eax
77e964cf push 0x0
77e964d1 55 push ebp
77e964d2 8bec mov ebp,esp
77e964d4 6aff push 0xff
77e9651d ff750c push dword ptr [ebp+0xc] <-- push the argument
77e96520 ff5508 call dword ptr [ebp+0x8] <-- DWORD WINAPI
77e96523 50 push eax
77e96524 e805000000 call KERNEL32!ExitThread (77e9652e)
77e96529 e923f10000 jmp KERNEL32!BaseThreadStart+0x81 (77ea5651)
Judjing from the examples above I think we can conclude that stack frames tracing can stop when either the return address or the old ebp turns to zero.
1.3. "Called from" functionality
Up to now we know how to create a callstack list (a list of addresses of functions). I would like to say a few words about the following question: Is it possible to implement a method that will allow the following C++ functionality?
Two additional questions arise here:
1.3.1 Do we need such a functionality?
This functionality could be considered as a new type of C++ runtime information. Since it is not implemented in the C++ standard it is probably useless... In my opinion this is a theoretical question and it should be the subject of a separate discussion.
1.3.2. How to implement this functionality?
Since we know the call stack of
it may seems trivial to browse the caller addresses searching for the address of
. Unfortunately things are not so simple. The main difficulty is how to get the address of
may be a class member (virtual or nonvirtual) function. There is no way to get the "real" address of a function from inside a C++ program. By "real address" I mean the relative virtual address where the compiler placed the function body. In most cases when you get a function address from inside a C++ program it does not appear to be a real address, but rather an address in some virtual or thunking table. On the other hand it is definitely clear that callstack addresses are real addresses. So what can we do in this case? The only simple solution I have found is to use function names instead of their addresses. This is because there is a relatively easy way(s) to find the address of a function, having given its name. This will allow us to implement the above function as follows:
if (IsCalledFrom("functionX") || IsCalledFrom("A::member1"))
Now the question is how to find the address of a function given its name. There are two common ways to do this:
- The development environment provides libraries that allow working with symbolic information. For example Microsoft provides the debug help library named
DbgHelp (prior to Windows 2000 the library was known as Image Help Library). The library contains functions for working with symbolic information, for example
SymFromName etc. As an example of using the
DbgHelp library here is the function
getFuncInfo you can find in the source. As far as I know, Borland provides a similar library named Borland Debug Hook Library, that can be used to extract information from Borland debug symbol (
.tds) files. Unfortunatelly there is not much information on the net about how to use this library. You can download it from here.
- Working with .map files. Both Microsoft VC++ and Borland C++ compilers are able to generate map files. Generally speaking map files are text files that contain information about functions (and variables) in a module and their addresses. For example here is a snippet of a
.map file generated by the Borland C++ 5.02 compiler:
0001:000006F5 StackDumper::DumpStack(unsigned int)
0001:000006C5 StackDumper::GetFuncAddr(unsigned long)
The important thing here is that the addresses are logical addresses. For example 0001:00000639 means that the destructor
resides in the first section in the PE file at offset 0x639 in that section. In the source you can find a simple function (written by Matt Pietrek) that converts linear addresses to logical addresses. Using this function you can convert the addresses from the stack trace into logical addresses that can be found in the map file. Having the logical address you just have to parse the
file in order to find the function name. As a conclusion I have to say that I don't have any idea about how to implement the
functionality in your release builds - when you have neither debug information nor a map file generated (or you don't want to distribute such information with your program).
2. Exit thunks
In this section I will describe how to use the stack frame information in order to implement exit thunks.
An exit thunk is a function that is invoked immediately after the
ret instruction of the function for which the thunk is installed. Exit thunks can be used for example in profiling applications. Here they are implemented as just another example of how to use the stack frame information.
The idea is very simple - in order to install an exit thunk we just have to declare a local variable:
StackDumper varName(true) (
true means "use exit thunk") in the body of the function for which we want to install the thunk. The destructor
StackDumper::~StackDumper() first saves the original return address of the function(i.e.
StackDumper is declared in a static local variable in
StackDumper. Afterwards it replaces the return address of
foo with the address of the
ExitThunk. This causes the
ret instruction of
foo to pass the control to the beginning of the
ExitThunk. (Note that the destructor is the correct place to perform this replacement. If we replace the function return address in the constructor (for instance) subsequent calls to the
DumpStack function will generate erroneous stack trace information).
This is not a common function call - the
ret instruction has popped the return address from the stack (which is now the address of
ExitThunk) and a jump is performed. So if
ExitThunk builds a standard stack frame this frame will not contain a return address. Another problem is that
ExitThunk has to be "invisible" - it should not touch the registers, especially
eax - where the function
foo has placed its return value (if any). If
ExitThunk is a standard function it will have something like this at the beginning(MSVC, Debug):
50: void main()
004010F2 55 push ebp
004010F3 8B EC mov ebp,esp
004010F5 6A FF push 0FFh
004010F7 68 F0 30 41 00 push offset $L49591 (004130f0)
004010FC 64 A1 00 00 00 00 mov eax,fs: <---Look here:
The eax register
is changed and
there is nothing
you can do!!!
00401102 50 push eax
00401103 64 89 25 00 00 00 00 mov dword ptr fs:,esp
You see that
could not be a normal function.This is the reason for which this function must be declared
. From MSDN: "For functions declared with the naked attribute, the compiler generates code without prolog and epilog code. You can use this feature to write your own prolog/epilog code using inline assembler code.". Fortunately the three compilers I have tested the examples with (BC++ 5.02, BCB6, MSVC6) support naked function calls. (For Borland C++ 5.02 users this is probably a surprise. This was not documented. ;-))).This is the solution to the above mentioned problems. So the implementation of
could be as follows:
void __declspec(naked) StackDumper::ExitThunk()
mov ebp, esp
sub esp, 4 pushad
temp = origRetAddr;
mov esi, temp
jmp esi -> can not use ret here.
Well, that's all folks!
I would like to thank mamaich for his help on disassembling issues!
Note about source compilation. In order to compile inline assembler source code with Borland C++ 5.02 you will need Turbo Assembler (
tasm32.exe is not included in the Borland C++ 5.02 distribution. If you have BC++ Builder you will find
tasm32.exe in the