[Buildroot] The C++11 issue

[Romain Naour <romain.naour@openwide.fr>]

Hi Thomas,

Le 21/03/2015 10:28, Thomas Petazzoni a ?crit :
> Hello,
> 
> A number of packages start to depend on C++11 features, which are not
> available in all toolchains: toolchains using old compiler versions
> (such as the Blackfin one), but also some older Crosstool-NG generated
> toolchains.
> 
> This is causing a number of build failures these days, such as:
> 
>         bfin |                  jsoncpp-1.6.0 | NOK | http://autobuild.buildroot.net/results/7fac9e4b3aa0746c860fd032a4b7b8663fb46fc8/
>         bfin |                  jsoncpp-1.6.0 | NOK | http://autobuild.buildroot.net/results/20b3f33e61286891a7804352c25326ea9faccfbb/
>         bfin |                  jsoncpp-1.6.0 | NOK | http://autobuild.buildroot.net/results/e482ba1d06c7fe2736ab3e79796297db8fbea187/
>          arm |                  jsoncpp-1.6.0 | NOK | http://autobuild.buildroot.net/results/1bf20d204042810bedad58dd2466bde9ae7b3fd6/
> 
> (std::snprintf is a C++11 feature)
> 
> So, it looks like we need to add a BR2_TOOLCHAIN_HAS_CXX11 option, with
> the usual handling.
> 
> However, it is not very clear-cut what is a C++11 supporting toolchain,
> and what is not: C++11 was added progressively in gcc, and there are
> still a few things missing apparently. According to
> https://gcc.gnu.org/projects/cxx0x.html, most of the compiler related
> features have been merged by gcc 4.8. But in terms of libstdc++
> support, the page
> https://gcc.gnu.org/onlinedocs/libstdc++/manual/status.html#status.iso.200x
> is a lot less clear, as it only indicates the mainline gcc status, and
> not the gcc releases.
> 
> So how do we decide whether a particular toolchain has C++11 support ?
> Do we simply decide that gcc >= 4.8 have sufficient C++11 support for
> now, and we'll adjust things later if needed ?
> 
> What do you think ?

Some host packages start to require c++11 too, like host-thrift.

http://autobuild.buildroot.net/?reason=host-thrift-0.9.2

(similar issue with target's thrift)
http://autobuild.buildroot.net/results/633/63340efff9adae966ec1d74ebc123d0437b75b63/build-end.log

In the case of host-thrift, the check for C++11 is horribly broken since there
is *sometime* some garbage in ax_cxx_compile_cxx11_required variable.

I can't reproduce the host-thrift issue on my Debian6 VM which use a gcc 4.4
because of that.

The custom m4 macro AX_CXX_COMPILE_STDCXX_11 used in configure.ac would stop the
build if no C++11 support is available:

AX_CXX_COMPILE_STDCXX_11([noext])

But if you look at the configure script generated by autoreconf,
ax_cxx_compile_cxx11_required is initialized with "truednl" !

ax_cxx_compile_cxx11_required=truednl

Then the build system fall back in C++11 optional requirement mode and finishes
the build...

All that to say that we might need also a check for C++11 for the host.

Otherwise, I like Arnout's proposal about BR2_GCC_AT_LEAST_4_X symbols.

Best regards,
Romain

> 
> Thomas
>