LiKao
LiKao

Reputation: 10658

Duplicate Destructors in Assembler Output for C++

I am currently trying to understand how G++ generates assembly from a small example C++ program. So I am testing the following programm with various optimizations:

#define noinline __attribute__ ((noinline))

class Test_Base {
public:
    noinline Test_Base() {}
    virtual noinline ~Test_Base() {}
    void noinline method() { v_method(); }
private:
    virtual void v_method()=0;
};

class Test1
   : public Test_Base
{
public:
    noinline Test1() {}
    noinline ~Test1() {}
private:
    void noinline v_method() {}
};

class Test2
   : public Test_Base
{
public:
    noinline Test2() {}
    ~Test2() {}
private:
    void noinline v_method() {}
};

int main() {
    volatile int x = 0;
    Test_Base * b;
    Test1 t1;
    Test2 t2;
    if( x )
        b = &t1;
    else
        b = &t2;
    b->method();
    return 0;
}

However looking at this code (compiled using -Os for the ARMv7 plattform), I found that all definitions of the constructors and destructors have been included multiple times. Here are the relevant parts of the symbol table for Test1:

00008730  w    F .text  00000004              Test1::v_method()
000088d8  w    O .rodata        00000014              vtable for Test1
000087d0  w    F .text  00000020              Test1::Test1()
00008774  w    F .text  0000001c              Test1::~Test1()
00008710  w    F .text  00000020              Test1::~Test1()
000088a4  w    O .rodata        00000007              typeinfo name for Test1
00008898  w    O .rodata        0000000c              typeinfo for Test1
000087d0  w    F .text  00000020              Test1::Test1()
00008710  w    F .text  00000020              Test1::~Test1()

So I have one constructor and two destructors (the last two calls are just duplicates tat the same positions as before). Looking at the assembly I observe the following:

First the constructor

000087d0 <Test1::Test1()>:
    87d0:       e92d4010        push    {r4, lr}
    87d4:       e1a04000        mov     r4, r0
    87d8:       ebfffff3        bl      87ac <Test_Base::Test_Base()>
    87dc:       e1a00004        mov     r0, r4
    87e0:       e59f3004        ldr     r3, [pc, #4]    ; 87ec <Test1::Test1()+0x1c>
    87e4:       e5843000        str     r3, [r4]
    87e8:       e8bd8010        pop     {r4, pc}
    87ec:       000088e0        .word   0x000088e0

I guess this does what I told it to do.

Now the destructor at 0x8710:

00008710 <Test1::~Test1()>:
    8710:       e59f3014        ldr     r3, [pc, #20]   ; 872c <Test1::~Test1()+0x1c>
    8714:       e92d4010        push    {r4, lr}
    8718:       e1a04000        mov     r4, r0
    871c:       e5803000        str     r3, [r0]
    8720:       ebfffff6        bl      8700 <Test_Base::~Test_Base()>
    8724:       e1a00004        mov     r0, r4
    8728:       e8bd8010        pop     {r4, pc}
    872c:       000088e0        .word   0x000088e0

Again there is nothing suspicious here.

Now the destructor at 0x8774:

00008774 <Test1::~Test1()>:
    8774:       e92d4010        push    {r4, lr}
    8778:       e1a04000        mov     r4, r0
    877c:       ebffffe3        bl      8710 <Test1::~Test1()>
    8780:       e1a00004        mov     r0, r4
    8784:       ebffff69        bl      8530 <_init+0x44>
    8788:       e1a00004        mov     r0, r4
    878c:       e8bd8010        pop     {r4, pc}

I cannot realy tell what this does, as I am not really familiar with the ABI. I am guessing it has something to do with static initialization.

What is the purpose of this additional destructor?

If I compile the same for x86_64 I also get duplicated destructors, so this does not seem system specific.

Upvotes: 2

Views: 172

Answers (1)

Mike Seymour
Mike Seymour

Reputation: 254501

The first is a non-virtual destructor, used for destroying automatic or static objects when the dynamic type is known at compile time.

The second is a virtual "thunk" used for polymorphic deletion. It calls the non-virtual destructor to destoy the object, then calls operator delete to free the memory.

Upvotes: 2

Related Questions