CXXABI_TM_1 status

CXXABI_TM_1 status

[Matthieu DARBOIS]

Hi,
a new manylinux policy has been defined for python.
This includes a series of allowed symbol versions from libstdc++ as
can be seen here
https://github.com/pypa/auditwheel/pull/192#discussion_r331555495

We'd like confirmation that the CXXABI_TM_1 symbol version can be
safely included in this policy.
From what I understand from reading the sources, I'd say that it's
safe to include this symbol version and that it won't disappear
overnight (seems part of the a stable ABI since gcc 4.7.0).

Could we get confirmation please ?

Re: CXXABI_TM_1 status

[Jonathan Wakely]

On Mon, 7 Oct 2019 at 22:13, Matthieu DARBOIS wrote:
>
> Hi,
> a new manylinux policy has been defined for python.
> This includes a series of allowed symbol versions from libstdc++ as
> can be seen here
> https://github.com/pypa/auditwheel/pull/192#discussion_r331555495
>
> We'd like confirmation that the CXXABI_TM_1 symbol version can be
> safely included in this policy.
> From what I understand from reading the sources, I'd say that it's
> safe to include this symbol version and that it won't disappear
> overnight (seems part of the a stable ABI since gcc 4.7.0).

Yes, it's part of the ABI baseline and can't be removed from
libstdc++.so without changing the SONAME of the library.

But why do you want to allow that symbol in manylinux binaries anyway?
Are people really using the experimental Transactional Memory
extensions in Python modules?

Re: CXXABI_TM_1 status

[Matthieu DARBOIS]

On Tue. 8 Oct. 2019 at 10:45, Jonathan Wakely <jwakely.gcc@gmail.com> wrote :
>
> On Mon, 7 Oct 2019 at 22:13, Matthieu DARBOIS wrote:
> >
> > Hi,
> > a new manylinux policy has been defined for python.
> > This includes a series of allowed symbol versions from libstdc++ as
> > can be seen here
> > https://github.com/pypa/auditwheel/pull/192#discussion_r331555495
> >
> > We'd like confirmation that the CXXABI_TM_1 symbol version can be
> > safely included in this policy.
> > From what I understand from reading the sources, I'd say that it's
> > safe to include this symbol version and that it won't disappear
> > overnight (seems part of the a stable ABI since gcc 4.7.0).
>
> Yes, it's part of the ABI baseline and can't be removed from
> libstdc++.so without changing the SONAME of the library.
>
> But why do you want to allow that symbol in manylinux binaries anyway?
> Are people really using the experimental Transactional Memory
> extensions in Python modules?

I'm not sure anyone will be using Transactional Memory extensions in
Python modules but since CXXABI_TM_1 symbol version is part of the ABI
baseline there's also no reason to forbid its usage. Is there ?