RFC: Drop C++98 support for versioned namespace (i.e. unstable ABI mode)

[Jonathan Wakely]

On Sat, 28 Aug 2021 at 11:48, Jonathan Wakely wrote:
>
> Not something I plan to work on in the near future, but an idea worth sharing.
>
> Before anybody panics, this does not affect the default configuration.
> C++98 will still be supported. I'm only talking about the unstable ABI
> build mode chosen when installing GCC, by
> --enable-symvers=gnu-versioned-namespace. If you don't know what that
> is, this RFC isn't something you need to worry about.
>
> Currently there are some places where we can't define constructors,
> assignment operators and destructors as defaulted, because it would
> create an ABI incompatibility between C++98 (which can't use defaulted
> members) and later standards.
>
> That means that even for the unstable ABI (versioned namespace)
> configuration, we are constrained by backwards compatibility (but
> compat with older standards, not older GCC releases).
>
> If we dropped C++98 support for the versioned namespace, it would open
> up additional opportunities for improvements to the versioned
> namespace mode.
>
> As a specific example, std::allocator currently has a user-provided,
> non-trivial default constructor. We can't define it as =default
> because that doesn't work in C++98, and would make it trivial in
> C++11, but still user-provided (and non-trivial) in C++98. We can't
> make that change for the ABI stable mode anyway, because it would be
> incompatible with older GCC releases. But we could do it for the
> versioned namespace if we stopped supporting C++98.

Maybe a more compelling example is that we could use
[[__no_unique_address__]] in _Rb_tree_compare to reduce the size of
std::map and std::set by alignof(void*) for the common case, see PR
102015. We can't do that because std::set and std::map need to compile
as C++98, and so can't use [[foo]] attributes, and there's no
__attribute__((__no_unique_address__)) alternative. So there's an
obvious improvement that would reduce object size from 48 bytes to 40
bytes, but we can't even do it for the ABI unstable mode because of
C++98 support.

I think anybody using the "I don't care about ABI compatibility, give
me the most optimized libstdc++ possible" mode but also expecting
C++98 support should reconsider their poor life choices.