Does C++ virtual function call on derived object g

问题:

In the following code, it calls a virtual function foo via a pointer to a derived object. Will this call go through the vtable or will it call B::foo directly?

If it goes via a vtable, what would be a C++ idiomatic way of making it call B::foo directly? I know that in this case I am always pointing to a B.

Class A
{
    public:
        virtual void foo() {}
};

class B : public A
{
    public:
        virtual void foo() {}
};


int main()
{
    B* b = new B();
    b->foo();
}

回答1:

Yes, it will use the vtable (only non-virtual methods bypass the vtable). To call B::foo() on b directly, call b->B::foo().

回答2:

Most compilers will be smart enough to eliminate the indirect call in that scenario, if you have optimization enabled. But only because you just created the object and the compiler knows the dynamic type; there may be situations when you know the dynamic type and the compiler doesn't.

回答3:

As usual, the answer to this question is "if it is important to you, take a look at the emitted code". This is what g++ produces with no optimisations selected:

18     b->foo();
0x401375 <main+49>:  mov    eax,DWORD PTR [esp+28]
0x401379 <main+53>:  mov    eax,DWORD PTR [eax]
0x40137b <main+55>:  mov    edx,DWORD PTR [eax]
0x40137d <main+57>:  mov    eax,DWORD PTR [esp+28]
0x401381 <main+61>:  mov    DWORD PTR [esp],eax
0x401384 <main+64>:  call   edx

which is using the vtable. A direct call, produced by code like:

B b;
b.foo();

looks like this:

0x401392 <main+78>:  lea    eax,[esp+24]
0x401396 <main+82>:  mov    DWORD PTR [esp],eax
0x401399 <main+85>:  call   0x40b2d4 <_ZN1B3fooEv>

回答4:

This is the compiled code from g++ (4.5) with -O3

_ZN1B3fooEv:
    rep
    ret

main:
    subq    $8, %rsp
    movl    $8, %edi
    call    _Znwm
    movq    $_ZTV1B+16, (%rax)
    movq    %rax, %rdi
    call    *_ZTV1B+16(%rip)
    xorl    %eax, %eax
    addq    $8, %rsp
    ret

_ZTV1B:
    .quad   0
    .quad   _ZTI1B
    .quad   _ZN1B3fooEv

The only optimization it did was that it knew which vtable to use (on the b object). Otherwise "call *_ZTV1B+16(%rip)" would have been "movq (%rax), %rax; call *(%rax)". So g++ is actually quite bad at optimizing virtual function calls.

回答5:

Compiler can optimize away virtual dispatch and call virtual function directly or inline it if it can prove it's the same behavior. In the provided example, compiler will easily throw away every line of code, so all you'll get is this:

int main() {}

回答6:

I changed the code up a bit to give it a go myself, and to me it looks like it's dropping the vtable, but I'm not expert enough in asm to tell. I'm sure some commentators will set me right though :)

struct A {
    virtual int foo() { return 1; }
};

struct B : public A {
    virtual int foo() { return 2; }
};

int useIt(A* a) {
    return a->foo();
}

int main()
{
    B* b = new B();
    return useIt(b);
}

I then converted this code to assembly like this:

g++ -g -S -O0  -fverbose-asm virt.cpp 
as -alhnd virt.s > virt.base.asm
g++ -g -S -O6  -fverbose-asm virt.cpp 
as -alhnd virt.s > virt.opt.asm

And the interesting bits look to me like the 'opt' version is dropping the vtable. It looks like it's creating the vtable but not using it..

In the opt asm:

9:virt.cpp      **** int useIt(A* a) { 
89                    .loc 1 9 0 
90                    .cfi_startproc 
91                .LVL2: 
10:virt.cpp      ****     return a->foo(); 
92                    .loc 1 10 0 
93 0000 488B07        movq    (%rdi), %rax    # a_1(D)->_vptr.A, a_1(D)->_vptr.A 
94 0003 488B00        movq    (%rax), %rax    # *D.2259_2, *D.2259_2 
95 0006 FFE0          jmp *%rax   # *D.2259_2 
96                .LVL3: 
97                    .cfi_endproc 

and the base.asm version of the same:

  9:virt.cpp      **** int useIt(A* a) { 
  88                    .loc 1 9 0 
  89                    .cfi_startproc 
  90 0000 55            pushq   %rbp    # 
  91                .LCFI6: 
  92                    .cfi_def_cfa_offset 16 
  93                    .cfi_offset 6, -16 
  94 0001 4889E5        movq    %rsp, %rbp  #, 
  95                .LCFI7: 
  96                    .cfi_def_cfa_register 6 
  97 0004 4883EC10      subq    $16, %rsp   #, 
  98 0008 48897DF8      movq    %rdi, -8(%rbp)  # a, a 
  10:virt.cpp      ****     return a->foo(); 
  99                    .loc 1 10 0 
 100 000c 488B45F8      movq    -8(%rbp), %rax  # a, tmp64 
 101 0010 488B00        movq    (%rax), %rax    # a_1(D)->_vptr.A, D.2263 
 102 0013 488B00        movq    (%rax), %rax    # *D.2263_2, D.2264 
 103 0016 488B55F8      movq    -8(%rbp), %rdx  # a, tmp65 
 104 001a 4889D7        movq    %rdx, %rdi  # tmp65, 
 105 001d FFD0          call    *%rax   # D.2264 
  11:virt.cpp      **** } 
 106                    .loc 1 11 0 
 107 001f C9            leave 
 108                .LCFI8: 
 109                    .cfi_def_cfa 7, 8 
 110 0020 C3            ret 
 111                    .cfi_endproc 

On line 93 we see in the comments: _vptr.A which I'm pretty sure means it's doing a vtable lookup, however, in the actual main function, it seems to be able to predict the answer and doesn't even call that useIt code:

 16:virt.cpp      ****     return useIt(b);
 17:virt.cpp      **** }
124                    .loc 1 17 0
125 0015 B8020000      movl    $2, %eax    #,

which I think is just saying, we know we're gonna return 2, lets just put it in eax. (I re ran the program asking it to return 200, and that line got updated as I would expect).

---

extra bit

So I complicated the program up a bit more:

struct A {
    int valA;
    A(int value) : valA(value) {}
    virtual int foo() { return valA; }
};

struct B : public A {
    int valB;
    B(int value) : valB(value), A(0) {}
    virtual int foo() { return valB; }
};

int useIt(A* a) {
    return a->foo();
}

int main()
{
    A* a = new A(100);
    B* b = new B(200);
    int valA = useIt(a);
    int valB = useIt(a);
    return valA + valB;
}

In this version, the useIt code definitely uses the vtable in the optimized assembly:

  13:virt.cpp      **** int useIt(A* a) {
  89                    .loc 1 13 0
  90                    .cfi_startproc
  91                .LVL2:
  14:virt.cpp      ****     return a->foo();
  92                    .loc 1 14 0
  93 0000 488B07        movq    (%rdi), %rax    # a_1(D)->_vptr.A, a_1(D)->_vptr.A
  94 0003 488B00        movq    (%rax), %rax    # *D.2274_2, *D.2274_2
  95 0006 FFE0          jmp *%rax   # *D.2274_2
  96                .LVL3:
  97                    .cfi_endproc

This time, the main function inlines a copy of useIt, but does actually do the vtable lookup.

---

What about c++11 and the 'final' keyword?

So I changed one line to:

virtual int foo() override final { return valB; }

and the compiler line to:

g++ -std=c++11 -g -S -O6  -fverbose-asm virt.cpp

Thinking that telling the compiler that it is a final override, would allow it to skip the vtable maybe.

Turns out it still uses the vtable.

---

So my theoretical answer would be:

• I don't think there are any explicit, "don't use the vtable" optimizations. (I searched through the g++ manpage for vtable and virt and the like and found nothing).
• But g++ with -O6, can do a lot of optimization on a simple program with obvious constants to the point where it can predict the result and skip the call altogether.
• However, once things get complex (read real) it's definitely doing vtable lookups, pretty much everytime you call a virtual function.