-Wno-non-virtual-dtor

[Stephan Bergmann]

On 03/13/2012 03:00 PM, Lubos Lunak wrote:
> On Tuesday 13 of March 2012, Stephan Bergmann wrote:
>> On 03/13/2012 11:43 AM, Tor Lillqvist wrote:
>>> Hmm, now that the reason for using  -Wno-non-virtual-dtor has been
>>> documented (760e0d2d7329ca6fc00a8439715bae38becb168a ), I wonder,
>>> should we globally then also turn off the corresponding MSVC warning?
>>>
>>> That would be kinda predictable (from a "the usual waste of time"
>>> point of view), as I and others have committed over times dozens of
>>> WaE fixes for this very issue... (I.e. added a virtual no-op
>>> destructor in most cases).
>>>
>>> Or does gcc and MSVC warn for different cases of lack of virtual
>>> destructor? Is it certain that in all cases this warning is bogus, in
>>> both the gcc and MSVC cases?
>>
>> I think the warning is generally non-bogus.  The problem is that we were
>> not able to change the cppumaker-generated C++ headers without breaking
>> backwards compatiblity[1], and GCC was somewhat over-ambitious with this
>> warning[2], so had to disable it -- even if we would have preferred to
>> keep it on.
>
>   The original message about the problem,
> http://markmail.org/message/664jsoqe6n6smy3b , mentions that it is not
> possible to selectively disable the warning for just the classes where we
> need to keep the binary compatibility.
>
>   Assuming that my attached testcase correctly matches the problem, I do not
> see a warning with either gcc-4.6.3 or clang-3.1r152540. So it looks like we
> can either just enable the warning and selectively disable it, or we can have
> a configure check if some older gcc version has the problem (well, given that
> #pragma diagnostic is recent with gcc, it'll need an #ifdef anyway).
>
> $ g++ -Wall dtor.cpp -Wnon-virtual-dtor -c

I'm going to add protected, non-virtual dtors to the problematic 
classes.  That's backward compatible (not 100%, in that it prevents 
deletes of objects through pointers to those non-virtual-dtor classes, 
but those are actually bugs, anyway) and clean, and no longer causes 
-Wnon-virtual-dtor warnings with recent GCCs.

With that in place, we can re-enable -Wnon-virtual-dtor in 
solenv/gbuild/platform/unxgcc.mk.  (I planned on bluntly always enabling 
it there, which might cause problems for --enable-werror on platforms 
with old GCC versions.  Basing enabling on a GCC version check or some 
configure check might indeed be better.)

Stephan