[Bug c++/65143] New: [C++11] missing devirtualization for virtual base in "final" classes

[Bug c++/65143] New: [C++11] missing devirtualization for virtual base in "final" classes

[balakrishnan.erode at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=65143

            Bug ID: 65143
           Summary: [C++11] missing devirtualization for virtual base in
                    "final" classes
           Product: gcc
           Version: unknown
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: c++
          Assignee: unassigned at gcc dot gnu.org
          Reporter: balakrishnan.erode at gmail dot com

When a class is marked final, it can devirtualize access to all base classes as
its layout is known. This is missing in gcc. 

struct A
{
  int i();
};

struct B : public virtual A
{
  int get()
  {
    return A::i() + 1;
  }
};

struct C final : public B
{
  int get()
  {
    return A::i() + 2;
  }
};

int foo(C& c)
{  
  return c.get(); // Need not go via vtable pointer as class C is final
}

int foo(B& b2)
{
  return b2.get(); // This has to go via vtable as most derived class can
change the location of A
}

Assembly: Both do virtual dispatch

foo(C&):
    subq    $8, %rsp
    movq    (%rdi), %rax
    addq    -24(%rax), %rdi
    call    one::A::i()
    addq    $8, %rsp
    addl    $2, %eax
    ret
foo(B&):
    subq    $8, %rsp
    movq    (%rdi), %rax
    addq    -24(%rax), %rdi
    call    one::A::i()
    addq    $8, %rsp
    addl    $1, %eax
    ret

Complete example:
gcc: http://goo.gl/U4KEvj
clang: http://goo.gl/PpQCkd  -- Clang does this optimization

[Bug c++/65143] [C++11] missing devirtualization for virtual base in "final" classes

[balakrishnan.erode at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=65143

Balakrishnan B <balakrishnan.erode at gmail dot com> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |balakrishnan.erode at gmail dot co
                   |                            |m
           See Also|                            |https://gcc.gnu.org/bugzill
                   |                            |a/show_bug.cgi?id=49488

--- Comment #1 from Balakrishnan B <balakrishnan.erode at gmail dot com> ---
I mean access to A::i() can be devirtualized in C::get().

[Bug c++/65143] [C++11] missing devirtualization for virtual base in "final" classes

[balakrishnan.erode at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=65143

--- Comment #2 from Balakrishnan B <balakrishnan.erode at gmail dot com> ---
Can someone please confirm the bug?

[Bug c++/65143] [C++11] missing devirtualization for virtual base in "final" classes

[maltsevm at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=65143

Mikhail Maltsev <maltsevm at gmail dot com> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |maltsevm at gmail dot com

--- Comment #3 from Mikhail Maltsev <maltsevm at gmail dot com> ---
This optimization is still missing in current trunk:

$ cat ./test.cc
struct A
{
    int j;
};

struct B : public virtual A
{
};

struct C final : public B
{
    int get();
};

int C::get()
{
    return A::j;
}

$ /opt/gcc-5.0.0/bin/g++ -O -std=c++11 -c -S ./test.cc -o test_gcc.s && cat
./test_gcc.s
        .file   "test.cc"
        .text
        .align 2
        .globl  _ZN1C3getEv
        .type   _ZN1C3getEv, @function
_ZN1C3getEv:
.LFB0:
        .cfi_startproc
        movq    (%rdi), %rax
        movq    -24(%rax), %rax
        movl    (%rdi,%rax), %eax
        ret
        .cfi_endproc
.LFE0:
        .size   _ZN1C3getEv, .-_ZN1C3getEv
        .ident  "GCC: (GNU) 5.0.0 20150315 (experimental)"
        .section        .note.GNU-stack,"",@progbits

(Note: optimization flags like -O3 and -fdevirtualize-speculatively don't
affect this behavior; neither method calls nor data member accesses get
devirtualized)

$ clang++ -O -std=c++11 -c -S ./test.cc -o test_clang.s && cat ./test_clang.s 
        .text
        .file   "./test.cc"
        .globl  _ZN1C3getEv
        .align  16, 0x90
        .type   _ZN1C3getEv,@function
_ZN1C3getEv:                            # @_ZN1C3getEv
        .cfi_startproc
# BB#0:                                 # %entry
        movl    8(%rdi), %eax
        retq
.Ltmp0:
        .size   _ZN1C3getEv, .Ltmp0-_ZN1C3getEv
        .cfi_endproc


        .ident  "clang version 3.7.0 (trunk 228487)"
        .section        ".note.GNU-stack","",@progbits

So, I can confirm it in sense "reproduce" (I don't have a right to change the
status of PR)

[Bug c++/65143] [C++11] missing devirtualization for virtual base in "final" classes

[balakrishnan.erode at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=65143

--- Comment #4 from Balakrishnan B <balakrishnan.erode at gmail dot com> ---
Thanks for confirming!

[Bug c++/65143] [C++11] missing devirtualization for virtual base in "final" classes

[trippels at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=65143

Markus Trippelsdorf <trippels at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |gpderetta at gmail dot com

--- Comment #5 from Markus Trippelsdorf <trippels at gcc dot gnu.org> ---
*** Bug 67184 has been marked as a duplicate of this bug. ***

[Bug c++/65143] [C++11] missing devirtualization for virtual base in "final" classes

[trippels at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=65143

Markus Trippelsdorf <trippels at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |trippels at gcc dot gnu.org

--- Comment #6 from Markus Trippelsdorf <trippels at gcc dot gnu.org> ---
Testcase from PR67184:

markus@x4 tmp % cat test.ii
struct V {
  virtual void foo();
};

struct wV final : V {};

struct oV final : V {
  void foo();
};

void call(V &x) { x.foo(); }

void call(wV &x) { x.foo(); }

void call(oV &x) { x.foo(); }

markus@x4 tmp % clang++ -std=c++14 -c -O3 -S test.ii -o -
...
_Z4callR2wV:                            # @_Z4callR2wV
        .cfi_startproc
# BB#0:                                 # %entry
        jmp     _ZN1V3fooEv             # TAILCALL
...

markus@x4 tmp % g++ -std=c++14 -O3 -c -S test.ii -o -
...
_Z4callR2wV:
.LFB1:
        .cfi_startproc
        movq    (%rdi), %rax
        jmp     *(%rax)

[Bug c++/65143] [C++11] missing devirtualization for virtual base in "final" classes

[paolo.carlini at oracle dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=65143

Paolo Carlini <paolo.carlini at oracle dot com> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|UNCONFIRMED                 |NEW
   Last reconfirmed|                            |2015-08-13
           See Also|                            |https://gcc.gnu.org/bugzill
                   |                            |a/show_bug.cgi?id=63164
     Ever confirmed|0                           |1

--- Comment #7 from Paolo Carlini <paolo.carlini at oracle dot com> ---
I'm not sure the two bugs are exactly equivalent. Anyway, when working on these
optimization issues, let' remember c++/63164 too.