C++ Logo

STD-PROPOSALS

Advanced search

Subject: Re: [std-proposals] Diamond dependencies and modules
From: Arthur O'Dwyer (arthur.j.odwyer_at_[hidden])
Date: 2020-01-22 09:39:35


It's a hard problem. Can I assume you've seen Titus Winters' talk "C++ as a
Live-at-Head Language"? If not, you definitely should.
https://www.youtube.com/watch?v=tISy7EJQPzI
He discusses this exact problem: the "dreaded diamond" as applied to
library dependencies. It has similar problems to the "dreaded diamond" in
class inheritance: If I have two copies of library T, then whose global
variables do I use?
Titus's conclusion IIRC is that there is no technical solution (especially
not Python-style semver), and the only practical solution is to "live at
head" and upgrade all your libraries frequently (and for all your suppliers
to upgrade *their* libraries — e.g. in your example it's really the
maintainer of library Y who's at fault, for making it impossible to use
Y-trunk with T-trunk).

The (well, one) problem with Jake's idea of "version namespaces" (and the
reason C++11's "inline namespaces" didn't help anything) is that namespaces
apply only to names, whereas ABI affects even things which are named by the
user, not the library. For example:

    struct MyClass {
        std::string s;
    };
    void foo(const MyClass&);

It doesn't matter whether `std::string` means `std::__1::string` or
`std::__2::string` — once it's wrapped in `MyClass`, you've lost any
benefit that name-mangling could possibly have given you. The name-mangling
of `MyClass` doesn't depend on the inline namespaces of its data members
(or the return types of its member functions, or anything else
ABI-significant like that).

–Arthur

On Wed, Jan 22, 2020 at 8:06 AM Jake Arkinstall via Std-Proposals <
std-proposals_at_[hidden]> wrote:

> In general this is something external to the standard, as it's more of a
> linking/package management issue rather than a language one.
>
> The ideal solution would be to have version data included in name
> mangling. It can be done manually with version namespaces. Automating it
> depends on which versioning system is being used by each project, and in
> general becomes an ecosystem problem, with which (IMO) the standard should
> maintain a clear separation in order to prevent unnecessary complexity in
> projects that don't need such functionality.
>
> On Wed, 22 Jan 2020, 12:56 Askar Safin via Std-Proposals, <
> std-proposals_at_[hidden]> wrote:
>
>> Hi. Consider this situation: program X depends on libraries Y and Z, Y
>> depends on library T 1.0 and Z depends on library T 2.0. Versions T 1.0 and
>> T 2.0 export same symbols (but they are implemented differently). As well
>> as I understand usual C++ package managers simply refuse to build X in this
>> situation. Can we somehow solve this problem? For example, using modules?
>>
>> Askar Safin
>> --
>> Std-Proposals mailing list
>> Std-Proposals_at_[hidden]
>> https://lists.isocpp.org/mailman/listinfo.cgi/std-proposals
>>
> --
> Std-Proposals mailing list
> Std-Proposals_at_[hidden]
> https://lists.isocpp.org/mailman/listinfo.cgi/std-proposals
>



STD-PROPOSALS list run by std-proposals-owner@lists.isocpp.org

Standard Proposals Archives on Google Groups