Re: ../../gdbsupport/new-op.cc:137:1: error: ‘void operator delete [](void*, std::size_t)’ is a usual (non-placement) deallocation function in C++14 (or with -fsized-deallocation) [-Werror=c++14-compat]

[Simon Marchi <simon.marchi@polymtl.ca>]

On 2022-03-02 12:22, Andrew Burgess wrote:
> Simon Marchi via Gdb-patches <gdb-patches@sourceware.org> writes:
> 
>> On 2022-03-02 11:30, Andrew Burgess wrote:
>>> * 周春明(日月) via Gdb-patches <gdb-patches@sourceware.org> [2022-03-01 15:48:47 +0800]:
>>>
>>>>
>>>> Hi GDB maintainers,
>>>>
>>>> I tried to build new GDB12, but encounter below error, anyone could tell me how to fix it? Thanks!
>>>>
>>>>  CXX new-op.o
>>>> ../../gdbsupport/new-op.cc:32:13: error: ‘void operator delete(void*, std::size_t)’ is a usual (non-placement) deallocation function in C++14 (or with -fsized-deallocation) [-Werror=c++14-compat]
>>>>  extern void operator delete (void *p, std::size_t) noexcept;
>>>>  ^
>>>> ../../gdbsupport/new-op.cc:33:13: error: ‘void operator delete [](void*, std::size_t)’ is a usual (non-placement) deallocation function in C++14 (or with -fsized-deallocation) [-Werror=c++14-compat]
>>>>  extern void operator delete[] (void *p, std::size_t) noexcept;
>>>>  ^
>>>> ../../gdbsupport/new-op.cc:119:1: error: ‘void operator delete(void*, std::size_t)’ is a usual (non-placement) deallocation function in C++14 (or with -fsized-deallocation) [-Werror=c++14-compat]
>>>>  operator delete (void *p, std::size_t) noexcept
>>>>  ^
>>>> ../../gdbsupport/new-op.cc:137:1: error: ‘void operator delete [](void*, std::size_t)’ is a usual (non-placement) deallocation function in C++14 (or with -fsized-deallocation) [-Werror=c++14-compat]
>>>>  operator delete[] (void *p, std::size_t) noexcept
>>>
>>> I was able to reproduce this on Ubuntu 16.04.1 with their gcc 5.4.0.
>>> I was unable to easily rebuild GCC 5.4.0 on my current development
>>> machine to check if this is reproducible with upstream gcc, or is just
>>> something impacting Ubuntu.  However, you can configure like:
>>>
>>>   ../src/configure ...configure-flags-here... CXXFLAGS="-Wno-error=c++14-compat"
>>>
>>> to disable this warning/error which I believe should be fine.
>>>
>>> For why this error is occurring, I'm honestly not 100% sure what the
>>> error is telling us.  I _think_ what it's saying is that the delete
>>> operator that we're declaring/defining conflicts with a "usual
>>> deallocation function", which is added in c++14 and means something
>>> specific.  I guess the idea is that maybe we're just randomly defining
>>> this version of delete for some reason, and then, if/when we move on
>>> to c++14 this function will get called unexpectedly by the language
>>> runtime in some situations.
>>>
>>> As time moved on I think this warning was relaxed, possibly with this
>>> commit:
>>>
>>>   https://gcc.gnu.org/legacy-ml/gcc-patches/2015-05/msg01883.html
>>>
>>> All this makes me wonder if the usual deallocation functions are ever
>>> actually used, and indeed, I applied the patch below, and GDB still
>>> seems to build fine, so this might be an alternative approach.  Maybe
>>> we should commit this to master?
>>>
>>> Thanks,
>>> Andrew
>>>
>>> ---
>>>
>>> diff --git a/gdbsupport/new-op.cc b/gdbsupport/new-op.cc
>>> index 1d066ba..4faa557 100644
>>> --- a/gdbsupport/new-op.cc
>>> +++ b/gdbsupport/new-op.cc
>>> @@ -27,11 +27,6 @@
>>>  #include "host-defs.h"
>>>  #include <new>
>>>  
>>> -/* These are declared in <new> starting C++14.  Add these here to enable
>>> -   compilation using C++11. */
>>> -extern void operator delete (void *p, std::size_t) noexcept;
>>> -extern void operator delete[] (void *p, std::size_t) noexcept;
>>
>> The story of this is that ASan gave some alloc-dealloc mismatch warnings
>> if we didn't define these specific delete operators, so we defined them
>> in this commit:
>>
>>   https://gitlab.com/gnutools/binutils-gdb/-/commit/5fff6115feae7aaa23c0ae8d144e1c8418ee2ee1
> 
> I saw that commit, but it wasn't clear (to me) that _every_ supplied
> delete was actually needed, or if the patch author just went wide.
> 
> The patch even says:
> 
>   There could be 16 operators delete but there are only 6, GDB uses 2 of
>   them.  It depends on libraries and compiler which of the operators
>   will get used.
> 
> But, that does seem to indicate that if I built with a different
> compiler/c++-runtime, then I might end up using different delete
> functions. 
> 
>>
>> But for the variants of delete that are only introduced in C++14, it
>> meant adding functions without an equivalent declaration when building
>> in C++11, which produced some -Wmissing-declarations warnings.  So these
>> declarations were added in this commit:
>>
>>   https://gitlab.com/gnutools/binutils-gdb/-/commit/b038b53f1ff4bf00ecdead1db23eddc4fd654305
> 
> OK, so, I think you're saying that if we compile with C++14 then we
> might need the delete functions I proposed deleting.
> 
> Just for my sanity, GDB currently targets C++11, right?  But, I guess
> you're suggesting you'd like to keep these functions in place so we
> _can_ compile with C++14 if we want?

That was my idea, yes.  We require GDB to be buildable in C++11, but we
also regularly build it in C++14 and C++17 as well, since recent
compilers default to that version (and we don't force
-std=c++11/-std=gnu++11).

>> The idea being that in C++11, these delete operators won't get used, but
>> there will be a declaration to avoid the warning, and in C++14 the
>> declarations will duplicate those found in headers, which is harmless.
>>
>> And now here we are, we have a C++11 compiler that complains about
>> declaring these delete operators.
>>
>> So we can't simply remove the declarations and / or definitions, we
>> would just re-introduce the problems fixed by these commits.
> 
> I think this last statement depends on the "wanting to compile with
> C++14" idea above, right? Otherwise...

Yes.

>> I think that a clean way to fix this would be to conditionally define
>> these delete operators based on the C++ version.  So we would remove the
>> declarations, as you do in your commit, but place the definitions under
>> an "#if __cplusplus >= xyz".
> 
> ... if we did this, and only ever compile with C++11, then this would be
> equivalent to what I proposed, right?

Yes (but we currently don't restrict to C++11).

> I ask the above not (just) to be pedantic, but (mostly) to ensure I've
> properly understood what's going on.
> 
> Below is a patch which maybe does what you suggest.  No commit message,
> but if this is what you were thinking then I can write this up.

I think it is what I suggested, by per Pedro's message, that is probably
not good.  From what I understand, it is possible for a library GDB
links with to be built with a more recent C++, and call some "recent"
delete operator.  So even if we build GDB in C++11, we have to cover the
more recent delete operators introduced in C++14.

Simon