[Bug c++/66701] New: __cxxabiv1::__cxa_pure_virtual - can it take an argument of the pointer to the function that was called, please?

[Bug c++/66701] New: __cxxabiv1::__cxa_pure_virtual - can it take an argument of the pointer to the function that was called, please?

[jmcguiness at liquidcapital dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=66701

            Bug ID: 66701
           Summary: __cxxabiv1::__cxa_pure_virtual - can it take an
                    argument of the pointer to the function that was
                    called, please?
           Product: gcc
           Version: 4.8.4
            Status: UNCONFIRMED
          Severity: enhancement
          Priority: P3
         Component: c++
          Assignee: unassigned at gcc dot gnu.org
          Reporter: jmcguiness at liquidcapital dot com
  Target Milestone: ---

I am aware of the ability to modify the handler for the "pure virtual function
called" error message via something like:

namespace __cxxabiv1 {
        extern "C" void
        __cxa_pure_virtual(void) {
                // Print a backtrace, etc...
                std::terminate();
        }
}

It is a great shame that the default, gcc-supplied function declaration does
not take an argument of the function pointer that was called, then the RTTI
information could be output. Note that the error message comes from a pure
virtual base class on which a pure virtual function has been called. Therefore
one would suppose that the was guaranteed to be RTTI information regarding that
function call, if not the enclosing pure virtual base class. So the default
implementation of __cxa_pure_virtual() would be able to call
"typeid(ptr_to_mem_fn).name()" to provide actual type information regarding the
class and the function that was attempted to be called.

I realise that the declaration could be difficult, because the new argument to
__cxa_pure_virtual(...) would need to be able to cope with both void and
non-void return types, cv-qualifiers, and arbitrary arguments. i.e. it would
need some form of declaration "magic". But the code is within the compiler! At
the call site where the canary is being placed, the type of the member function
is explicitly known, so the specific declaration (in particular the
ptr-to-mem-fn argument) of __cxa_pure_virtual() can be generated for that call
site, and therefore typeid(of the ptr-to-mem-fn arg) could be correctly
generated.

If I knew how g++ worked internally, I'd do this change. But I don't, I'm
sorry.

The issue I face is that I have a multi-threaded program that has a race
condition that causes the "pure virtual function call" error very rarely. The
code-base is very large, so it is very hard (I have tried) to track down where
the bug is. It does not appear in debug builds (-O0) and using my own
implementation of __cxa_pure_virtual() and calling set_terminate with my own
functions that generate backtraces has not created the backtrace I'd hope for.

I know also that the code I am referring to above is buggy, therefore I should
"just not write it". True. But surely improving the quality of the error
message with the class name and function name in the "pure virtual method
called" message would be og much help to the C++ community at large?

[Bug c++/66701] __cxxabiv1::__cxa_pure_virtual - can it take an argument of the pointer to the function that was called, please?

[redi at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=66701

Jonathan Wakely <redi at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|UNCONFIRMED                 |RESOLVED
         Resolution|---                         |INVALID

--- Comment #1 from Jonathan Wakely <redi at gcc dot gnu.org> ---
The function signature is defined by the ABI so we can't change it.

http://mentorembedded.github.io/cxx-abi/abi.html#pure-virtual

[Bug c++/66701] __cxxabiv1::__cxa_pure_virtual - can it take an argument of the pointer to the function that was called, please?

[gcc-bugs at hussar dot demon.co.uk]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=66701

--- Comment #2 from Jason <gcc-bugs at hussar dot demon.co.uk> ---
(With due sense of dread:) I note that the __cxxabiv1 seems to have a v1 in it.
Why not define a v2 with a better signature? Although I realise that this is a
quality-of-implementation issue, why should an apparently mis-declared function
signature be considered acceptable? I do not agree that this enhancement should
be marked as resolved invalid: I think that is a cop-out frankly.

[Bug c++/66701] __cxxabiv1::__cxa_pure_virtual - can it take an argument of the pointer to the function that was called, please?

[redi at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=66701

--- Comment #3 from Jonathan Wakely <redi at gcc dot gnu.org> ---
But this isn't the place to propose changes to the ABI, because it's used by
several different compilers and if G++ started emitting a call to a different
version of __cxa_pure_virtual that would break compatibility with other
compilers using the same ABI.

So this is not something that can be added to GCC unless it is first added to
the ABI.

[Bug c++/66701] __cxxabiv1::__cxa_pure_virtual - can it take an argument of the pointer to the function that was called, please?

[jmcguiness at liquidcapital dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=66701

Jason McG <jmcguiness at liquidcapital dot com> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|RESOLVED                    |CLOSED
         Resolution|INVALID                     |WONTFIX

--- Comment #4 from Jason McG <jmcguiness at liquidcapital dot com> ---
Once the existence of the restricted ABI was pointed out to me I did not
propose to change it. I proposed creating a new version of the ABI that would
have an improved signature. That signature would be dependent upon what the
compiler could emit. It is not unreasonable to assume that there is an
interplay between what the compiler can emit which should drive the
specification of the ABI.

I have edited the status to "closed/wontfix" as clearly there is no inclination
to implement this enhancement.

[Bug c++/66701] __cxxabiv1::__cxa_pure_virtual - can it take an argument of the pointer to the function that was called, please?

[jmcguiness at liquidcapital dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=66701

Jason McG <jmcguiness at liquidcapital dot com> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|CLOSED                      |UNCONFIRMED
         Resolution|WONTFIX                     |---

--- Comment #5 from Jason McG <jmcguiness at liquidcapital dot com> ---
Hang on: what has the ABI got to do with this? This discussion has got side
tracked onto a red herring. The code emitted by gcc apparently *eventually*
calls __cxa_pure_virtual, as we know. The signature of __cxa_pure_virtual could
be irrelevant: what is relevant is the *code path* that calls this function.
The code path could be enhanced to provide the extra information to the user.
The ABI function can then be left as is.

So I've re-opened the enhancement request.