DEV Community

Pierre Gradot
Pierre Gradot

Posted on • Edited on

vtable under the surface | Episode 3 - How virtual functions are actually called

In this episode, we will see how invoking a virtual function in C++ translates into assembly instructions. We will see how our class instance is constructed and how it relates to the vtable. Then, we will see how this vtable is used to call the appropriate function.

In you have actually built the project and analyzed the binary in the previous episode, don't forget to remove the -fno-rtti option and to rebuild the project. I will use this binary as a reference here.

Execute the Program with gdb

To understand how vtables are used in assembly, disassembling the binary with objdump --dissassemble could have been an option, but instead, I decided to use gdb to actually execute the program:

$ gdb a.out
Enter fullscreen mode Exit fullscreen mode

By default, gdb uses AT&T syntax to disassemble the code, but I prefer Intel flavor:

(gdb) set disassembly-flavor intel
(gdb) show disassembly-flavor
The disassembly flavor is "intel".

Enter fullscreen mode Exit fullscreen mode

Enabling name demangling will make it easier to understand the symbols being manipulated:

(gdb) set print asm-demangle
(gdb) show print asm-demangle
Demangling of C++/ObjC names in disassembly listings is on.
Enter fullscreen mode Exit fullscreen mode

Run to the main:

(gdb) break main
Breakpoint 1 at 0x1139: file /home/pierre/CLionProjects/untitled/main.cpp, line 3.
(gdb) run
Starting program: /home/pierre/CLionProjects/untitled/cmake-build-debug/a.out 
[Thread debugging using libthread_db enabled]
Using host libthread_db library "/lib/x86_64-linux-gnu/libthread_db.so.1".

Breakpoint 1, main () at /home/pierre/CLionProjects/untitled/main.cpp:3
3       int main() {
Enter fullscreen mode Exit fullscreen mode

We can now disassemble the code:

(gdb) disas
Dump of assembler code for function main():
=> 0x0000555555555139 <+0>:     sub    rsp,0x18
   0x000055555555513d <+4>:     mov    DWORD PTR [rsp+0x8],0x0
   0x0000555555555145 <+12>:    mov    DWORD PTR [rsp+0xc],0x0
   0x000055555555514d <+20>:    lea    rax,[rip+0x2c44]        # 0x555555557d98 <vtable for Derived+16>
   0x0000555555555154 <+27>:    mov    QWORD PTR [rsp],rax
   0x0000555555555158 <+31>:    mov    rdi,rsp
   0x000055555555515b <+34>:    call   0x5555555551c1 <use(Base const&)>
   0x0000555555555160 <+39>:    mov    eax,0x0
   0x0000555555555165 <+44>:    add    rsp,0x18
   0x0000555555555169 <+48>:    ret
End of assembler dump.
Enter fullscreen mode Exit fullscreen mode

We can clearly see the call to use(). The immediately preceding line prepares the first and only argument of this function (it puts it in the rdi register). The other lines above are the initialisation of obj, our instance of Derived. We can't understand how use() is called and uses the vtable to call Derived::foo() if we don't understand how obj is initialized and how it is connected to the vtable.

Clarification About Addresses

Before we dive in a line-by-line assembly analysis, I want to clarify why the address of use() in the call instruction is 0x005555555551c1 and not 0x00000000000011c1 (as in the symbol table from the previous episode, or in the disassembly produced by obdjump --disassemble).

When the binary is loaded in memory and executed, it's not placed at address 0x0000000000000000 but somewhere else. The addresses in the ELF file are relative from this "somewhere else".

We can compute this offset with a simple subtraction: 0x005555555551c1 - 0x00000000000011c1 = 0x00555555554000. The same translation can be applied to any other symbols. We can verify this offset with the address of main. We need the address of main in the binary:

$  objdump --syms a.out | grep 'main' | grep '.text'
0000000000001139 g     F .text  0000000000000031              main
Enter fullscreen mode Exit fullscreen mode

If we add the offset to this address, the result is the first address in the disassembly above: 0x0000000000001139 + 0x00555555554000 = 0x0000555555555139.

Construction of the Object

In main(), obj is initialized by these instructions (note that => indicates the line where gdb is paused):

=> 0x0000555555555139 <+0>:     sub    rsp,0x18
   0x000055555555513d <+4>:     mov    DWORD PTR [rsp+0x8],0x0
   0x0000555555555145 <+12>:    mov    DWORD PTR [rsp+0xc],0x0
   0x000055555555514d <+20>:    lea    rax,[rip+0x2c44]        # 0x555555557d98 <vtable for Derived+16>
   0x0000555555555154 <+27>:    mov    QWORD PTR [rsp],rax
Enter fullscreen mode Exit fullscreen mode

The first line moves the stack pointer, allocating space to store obj. We can execute a few commands to verify this:

(gdb) info reg rsp
rsp            0x7fffffffddd8      0x7fffffffddd8
(gdb) nexti
4           auto obj = Derived();
(gdb) info reg rsp
rsp            0x7fffffffddc0      0x7fffffffddc0
(gdb) print &obj
$2 = (Derived *) 0x7fffffffddc0
Enter fullscreen mode Exit fullscreen mode

After these commands, gdb is paused on the second line, at the address 0x000055555555513d. The address of obj is the same as the value of the rsp register. At this moment, the object is created but not initialized:

(gdb) print obj
$3 = {<Base> = {_vptr.Base = 0x0, dummy_base = 1431671456}, dummy_derived = 21845}
Enter fullscreen mode Exit fullscreen mode

It's not surprising to see dummy_base and dummy_derived, as they are the member data of Derived. But what is _vptr.Base? This is the "virtual pointer". It's automatically inserted by the compiler to reference the vtable. It doesn't point to (the beginning of) the vtable actually, but to a location that the ABI's specification refers to as the "virtual table address point", inside the vtable. This point corresponds to the first function pointer in the vtable. We will get back to this pointer later.

The next instructions are here to initialize obj.

First, dummy_base and dummy_derived are set to 0 by the two mov instructions (yes, even though our code doesn't require to initialize them). A C++ equivalent of mov DWORD PTR [rsp+0x8],0x0 would ressemble something like *(rsp + 0x8) = 0x0. At this point, the value of rsp is 0x007fffffffddc0, confirming that rsp+0x8 and rsp+0xc are indeed the addresses of the data members:

(gdb) print &obj.dummy_base
$5 = (int *) 0x7fffffffddc8
(gdb) print &obj.dummy_derived
$6 = (int *) 0x7fffffffddcc
Enter fullscreen mode Exit fullscreen mode

The initialization of the virtual pointer is slightly more complex:

   0x000055555555514d <+20>:    lea    rax,[rip+0x2c44]        # 0x555555557d98 <vtable for Derived+16>
   0x0000555555555154 <+27>:    mov    QWORD PTR [rsp],rax
Enter fullscreen mode Exit fullscreen mode

The lea instruction with an operand that is relative to the rip register is classic compiler technique to get the address of something that is inside the binary being executed. gdb kindly shows the result of the computation. It even tells us that this is the address of vtable for Derived+16. In the previous episode, we noted that the first function address in the vtable comes after 8 bytes for the offset and 8 bytes for the pointer to the RTTI, hence a total of 16 bytes. You should understand now why the virtual pointer is set to vtable for Derived+16: this is the "virtual table address point".

This address is temporarily stored into rax before being moved on the stack to complete the initialization of obj.

We can execute these instructions with gdb to run the initialization:

(gdb) ni
(gdb) ni
(gdb) ni
(gdb) ni
(gdb) print obj
$7 = {<Base> = {_vptr.Base = 0x555555557d98 <vtable for Derived+16>, dummy_base = 0}, dummy_derived = 0}
Enter fullscreen mode Exit fullscreen mode

If we want to verify what is being pointed to by the virtual pointer, we have 2 options.

The first solution is the manual, tedious one. We can dump the memory located at this address and compare the dumped values to the addresses of the member functions of Derived. We need to dump 2 sets of 8 bytes each (since a function pointer is 8-byte wide):

(gdb) x/2gx 0x555555557d98
0x555555557d98 <vtable for Derived+16>: 0x0000555555555196      0x00005555555551ac
(gdb) print 'Derived::foo'
$8 = {void (const Derived * const)} 0x555555555196 <Derived::foo() const>
(gdb) print 'Derived::bar'
$9 = {void (const Derived * const)} 0x5555555551ac <Derived::bar() const>
Enter fullscreen mode Exit fullscreen mode

The second solution is just using the dedicated command provided by gdb:

(gdb) info vtbl obj
vtable for 'Derived' @ 0x555555557d98 (subobject @ 0x7fffffffddc0):
[0]: 0x555555555196 <Derived::foo() const>
[1]: 0x5555555551ac <Derived::bar() const>
Enter fullscreen mode Exit fullscreen mode

Calling use() from main()

The object is now ready, and gdb is paused here:

(gdb) disas
(...)
=> 0x0000555555555158 <+31>:    mov    rdi,rsp
   0x000055555555515b <+34>:    call   0x5555555551c1 <use(Base const&)>
(...)
End of assembler dump.
Enter fullscreen mode Exit fullscreen mode

On Linux-x64, the calling convention for functions uses the rdi register as the first parameter. We've seen in the previous section that rsp holds the address of obj at this point. This address is moved in the rdi register and the function is ready to be called. Indeed, a reference in C++ is often just an address in assembly.

How the Virtual Function is Actually Called

We can reach the beginning of use() with ni (to execute the mov) and si (to step into the call). We can disassemble the function:

(gdb) disas
Dump of assembler code for function _Z3useRK4Base:
=> 0x00005555555551c1 <+0>:     sub    rsp,0x8
   0x00005555555551c5 <+4>:     mov    rax,QWORD PTR [rdi]
   0x00005555555551c8 <+7>:     call   QWORD PTR [rax]
   0x00005555555551ca <+9>:     add    rsp,0x8
   0x00005555555551ce <+13>:    ret
End of assembler dump.
Enter fullscreen mode Exit fullscreen mode

The virtual call is right here!

We have seen that the value of rdi is the virtual pointer and that the first entry in the array pointed to by this virtual pointer is the address of Derived::foo(). The value pointed to by rdi is moved into rax, and then the value pointed to by rax is called. The function Derived::foo() executes and prints Derived => foo()!

You can try to change the code to call bar() instead of foo(). The call instruction will be different. You will have call QWORD PTR [rax+0x8] instead, to get the next pointer in the array of function pointers.

Indirect Call

In use(), the instruction call QWORD PTR [rax] is an indirect call because the value of a register is used. The instruction call 0x5555555551c1 to call use() from main() is a direct call. The mnemonic is the same (call) but the opcodes (the binary values used to translate the mnemonic) are different.

We can examine the memory with gdb to see the difference:

(gdb) x /1bx use+7
0x5555555551c8 <use(Base const&)+7>:    0xff
(gdb) x /1bx main+34
0x55555555515b <main()+34>:     0xe8
Enter fullscreen mode Exit fullscreen mode

Indirect calls are typically used to implement function pointer.

Conclusion

In this episode, we discovered that the compiler adds a hidden field for each instance of a class with virtual functions. This field is the "virtual pointer" (or "vptr" in short). Thanks to this pointer, we can get a function pointer inside the vtable, and the desired function.

I will publish the episodes as I write them. Subscribe to be notified! 💃🏽

Top comments (0)