[Bug c++/64346] New: gcc generates incorrect debug info for ctor/dtor

[Bug c++/64346] New: gcc generates incorrect debug info for ctor/dtor

[dehao at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=64346

            Bug ID: 64346
           Summary: gcc generates incorrect debug info for ctor/dtor
           Product: gcc
           Version: 4.9.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: c++
          Assignee: unassigned at gcc dot gnu.org
          Reporter: dehao at gcc dot gnu.org

#cat a.cc
#include <iostream>
#include <memory>

class ABC {
public:
 ABC() {printf("abc\n");}
 ~ABC() {printf("xyz\n");}
};

int foo() {
  std::unique_ptr<ABC> a(new ABC());
}

#g++4.8 a.cc -c -O2 -std=c++11 -g -fno-exceptions
#addr2line -e a.o -i -f 0x13
_ZN3ABCC1Ev
/usr/local/google/home/dehao/gcc/gcc-4_9/debug/a.cc:5
_Z3foov
/usr/local/google/home/dehao/gcc/gcc-4_9/debug/a.cc:10

#g++4.9 a.cc -c -O2 -std=c++11 -g -fno-exceptions
#addr2line -e a.o -i -f 0x13
_ZN3ABCC4Ev
/usr/local/google/home/dehao/gcc/gcc-4_9/debug/a.cc:5
_Z3foov
/usr/local/google/home/dehao/gcc/gcc-4_9/debug/a.cc:10

#g++4.9 a.cc -c -O0 -std=c++11 -g -fno-exceptions
#nm a.o |grep ABCC
0000000000000000 W _ZN3ABCC1Ev
0000000000000000 W _ZN3ABCC2Ev
0000000000000000 n _ZN3ABCC5Ev

Note that for the gcc 4.9 version and later, without any optimization
_ZN3ABCC4Ev does not exists in the symbol table. So we expect either
_ZN3ABCC1Ev or _ZN3ABCC2Ev in the addr2line result.

[Bug c++/64346] gcc generates incorrect debug info for ctor/dtor

[dehao at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=64346

dehao at gcc dot gnu.org changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
           Priority|P3                          |P2
                 CC|                            |ccoutant at gcc dot gnu.org,
                   |                            |davidxl at google dot com,
                   |                            |jason at gcc dot gnu.org

--- Comment #1 from dehao at gcc dot gnu.org ---
This may be related to https://gcc.gnu.org/ml/gcc-patches/2013-11/msg02724.html

Jason could you help take a look?

[Bug c++/64346] gcc generates incorrect debug info for ctor/dtor

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=64346

--- Comment #2 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
>_ZN3ABCC4Ev

That is the internal name of the function (it demangles to ABC::ABC() anyways)
and the symbol does not need to be in the symbol table at all.

We are going to run into this more with ICF also.  I think there might have
been a RFC about how to handle the debugging for commoning of functions but it
might be a while before that is implemented.

[Bug c++/64346] gcc generates incorrect debug info for ctor/dtor

[dehao at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=64346

--- Comment #3 from dehao at gcc dot gnu.org ---
For AutoFDO, we actually needs symbols from the symbol table because indirect
call promotion needs the symbol name to find the right callee.

You are right, ICF also causes trouble to AutoFDO profile.

[Bug c++/64346] gcc generates incorrect debug info for ctor/dtor

[ccoutant at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=64346

--- Comment #4 from Cary Coutant <ccoutant at gcc dot gnu.org> ---
I think the problem is that the decloning turns the C1 and C2 ctors into
thunks, and GCC doesn't (yet) generate debug info for thunks, so when you
inline the C4 into the C1 or C2, and inline that into foo, all that's left in
the debug info is the inlined call to the C4 ctor.

It looks like -fno-declone-ctors-dtor doesn't work around the problem because
the "need_alias" flag is true, and that forces the decloning in this case.

The most expedient approach might be for you to alias the C4/D4 names to the
official ones when processing the profile data. Is that feasible?