llvm dev - Jul 2015 - [LLVMdev] Purpose of LLVM_ENABLE_LIBCXX and LLVM_ENABLE_LIBCXXABI

On 28 July 2015 at 14:44, Martell Malone <martellmalone at gmail.com>
wrote:
> Adding a little to the topic what criteria would we need to make a target
> use compiler-rt and libc++ as the default in the clang driver.
> I have successfully built a standalone clang toolchain with mingw-w64
> without using gcc or binutils.
There was another discussion with David Chisnall, where he gave a good
overall on some of the complications for doing so:

http://comments.gmane.org/gmane.comp.compilers.clang.devel/44037
>From my point of view, we can only do something that radical when we
can test on most environments and make sure that any combination will
work. I don't know how to do that, though, but I expect that there
should be enough tests in Clang to that effect.

For now, I just want to make sure it works on ARM/AArch64, even if I
have to hack my way through it. If I can have at least a buildbot
testing them, we'll be sure not to regress in functionality. System
integration is a secondary issue.

cheers,
--renato

Wouldn’t it it be possible to add a flag like the one used in the OpenMP
subproject? There they default to linking to the non-functional libgomp (as far
as I understand it), and if people want to use the actual libomp runtime, they
have to specify -fopenmp=libomp when compiling/linking a program. At the same
time, it is possible to use the libomp runtime library by default by specifying

-DCLANG_DEFAULT_OPENMP_RUNTIME=libomp

at configuration time when building LLVM. That way things would not break for
existing users with unsupported systems while making things for users like us
*much* easier :)

Regards,

Michael
> On 28 Jul 2015, at 15:53 , Renato Golin <renato.golin at linaro.org>
wrote:
> 
> On 28 July 2015 at 14:44, Martell Malone <martellmalone at gmail.com>
wrote:
>> Adding a little to the topic what criteria would we need to make a
target
>> use compiler-rt and libc++ as the default in the clang driver.
>> I have successfully built a standalone clang toolchain with mingw-w64
>> without using gcc or binutils.
> 
> There was another discussion with David Chisnall, where he gave a good
> overall on some of the complications for doing so:
> 
> http://comments.gmane.org/gmane.comp.compilers.clang.devel/44037
> 
> From my point of view, we can only do something that radical when we
> can test on most environments and make sure that any combination will
> work. I don't know how to do that, though, but I expect that there
> should be enough tests in Clang to that effect.
> 
> For now, I just want to make sure it works on ARM/AArch64, even if I
> have to hack my way through it. If I can have at least a buildbot
> testing them, we'll be sure not to regress in functionality. System
> integration is a secondary issue.
> 
> cheers,
> --renato

On 28 July 2015 at 16:57, Schlottke-Lakemper, Michael
<m.schlottke-lakemper at aia.rwth-aachen.de>
wrote:
> -DCLANG_DEFAULT_OPENMP_RUNTIME=libomp
>
> at configuration time when building LLVM. That way things would not break
for existing users with unsupported systems while making things for users like
us *much* easier :)
This is possible, but not yet done for compiler-rt / libc++. It needs
support in the driver to actually change the default. Shouldn't be too
hard, but we need something that can grow into architectural options
as well (like GCC), to make the life of distributions easier, so we
can't just hack up something that will work "just" for rt/libc++.

cheers,
--renato