[Date Prev][Date Next] [Thread Prev][Thread Next] [Date Index] [Thread Index]

Re: Mandates explicit -std=c++XY for c++ projects



On 10 October 2017 at 07:45, Mathieu Malaterre <malat@debian.org> wrote:
> Dear all,
>
> Since the GCC 6 release [1], the default mode for C++ is now
> -std=gnu++14 instead of -std=gnu++98. What this means is that upon
> (re)compilation a library written for c++98 will be recompiled using a
> different c++ standard (c++14 in this case), unless of course the
> upstream package explicitly set the -std= flags with the appropriate
> c++ version.
>
> The ISO committee generally describe the change in between different
> standards [2] and in some case, one can find examples of subtle change
> in behaviors [3] and [4].
>
> With this mind I'd like to make mandatory the -std=c++XY flags when
> compiling either a c++ library or a stand-alone c++ program:
>
> 1. Either upstream define the explicit -std=c++XY flags by mean of its
> build system,
> 2. Or the package maintainers needs to explicit change the CXXFLAGS to
> pass the appropriate version of the c++ standard. In which case this
> should be documented in the README.Debian file.
> 3. As a fallback, dh should initialize the CXXFLAGS with -std=gnu++98
>
> If there is a consensus on the following change, I'll go ahead and
> also file a bug for lintian to scan the compilation logs in search for
> missing -std=c++ expression when g++ command line are issued.

I wholeheartedly object. A given debian release has a defined default
ABI with a minimum CPU requirements per architecture. Currently, in
sid, it is C++11 abi. The way ABI works is that it forward compatible,
but not backwards. Therefore generally binaries compiled on older
releases are copied forward to newer releases and work fine (as long
as the depencies are still satisfiable, sans ocational bugs). Note
there are ABI fixes too, such that something compiled with c++98
standard in sid, is not guaranteed to have the same code generated nor
work with older runtime (e.g. old-stable). It does work for simplistic
cases however.

If you target and older release, and need a backport of a newer
software package it not acceptable to demand lowing std version in sid
such that packages are instalabble in older releases. As a
distribution, we cannot guarantee that to work.

In general, software should be written portable, and the default
standards version should be in use which is currently dictated by the
default compiler and thus maintainers of the default compilers in
Debian.

Specifically, there are many software packages in the archive that
compile correctly with both newer and older standards, and do
configure/compile time checking and take advantange of the newly
available standards features, instead of using embedded copies of code
/ adding additional libraries. There are many packages that use c++11
features, instead of including and linking boost libraries. Similarly
for C, there are packages that use standards available
alignment/padding/atomic functions, instead of using hand-crafted
embedded (often buggy in corner cases). Overall, it leads to software
packages reusing more shared code, as more things are available in the
standards (or streamlined in standards).

boost will continue to compile with the default C++ standards version,
and thus reverse-dependencies are also expected to use the matching /
default standards version. I have been pondering to provide stdc++98
and stdc++11 abi builds of boost, but that is a lot of code
duplication for effectively diminishing returns over time.

Encoding stdc++ version in the packages, will make them
non-binNMU-safe and thus will start to require sourcefull upload for
boost transitions which is a lot of packages.

If you have any examples of things failing to build due to change from
98 to 11 in unstable, I'd rather volunteer to fix that, rather than
hardcode the world to use old standards.

Using lowest common denominator will results in inefficient, large,
under-optimised binaries. One really should recompile software for
each distribution release toolchain.

-- 
Regards,

Dimitri.


Reply to: