[LLVMdev] 3.4 branch gcc 4.9 build error

[Tuncer Ayaz]

On 4/29/14, Markus Trippelsdorf wrote:
> On 2014.04.29 at 12:16 +0200, Tuncer Ayaz wrote:
> > On 4/29/14, Markus Trippelsdorf wrote:
> > > On 2014.04.29 at 11:04 +0200, Tuncer Ayaz wrote:
> > > > On 4/27/14, Markus Trippelsdorf wrote:
> > > > > On 2014.04.27 at 16:18 +0200, Tuncer Ayaz wrote:
> > > > > > Using gcc 4.9 to build the llvm 3.4 branch fails with the
> > > > > > following error:
> > > > > >
> > > > > > llvm[1]: Compiling APFloat.cpp for Release build
> > > > > > [...]
> > > > > > /usr/[...]/include/c++/4.9.0/cstddef:51:11:
> > > > > > error: no member named 'max_align_t' in the global namespace
> > > > > >   using ::max_align_t;
> > > > > >         ~~^
> > > > > >
> > > > > > gcc 4.8.2's caused no such issue, so it must be a change
> > > > > > in 4.9.0/cstddef.
> > > > > >
> > > > > > To be clear, I have no idea who's at fault (gcc 4.9.0
> > > > > > header or llvm sources), assuming any code is to be blamed
> > > > > > (at all).
> > > > >
> > > > > You can simply wrap it with #ifndef:
> > > > >
> > > > >  namespace std
> > > > >  {
> > > > >    // We handle size_t, ptrdiff_t, and nullptr_t in c++config.h.
> > > > > #ifndef __clang__
> > > > >    using ::max_align_t;
> > > > > #endif
> > > > >  }
> > > > >  #endif
> > > >
> > > > Yeah, but are you sure the bug is in libstdc++ and has to be
> > > > fixed in gcc upstream? I mean, would it be the wrong choice to
> > > > include a fix in the next llvm 3.4 release?
> > >
> > > One could argue that nothing needs to fixed. The libstdc++
> > > developers don't care very much about clang interoperability.
> > > The clang developers don't care about libstdc++ issues. Just use
> > > libc++ with clang and you're set.
> >
> > Maybe I wasn't clear, but the issue is building (bootstrapping)
> > llvm on a system with gcc-4.9 as the host compiler. Are you
> > suggesting that to build llvm with libstdc++-4.9 I should first
> > build libc++ with gcc-4.9, and use that lib (instead) when
> > building llvm? Is that even possible?
>
> No. During bootstrap you'll have to patch libstdc++. You can build
> and use libc++ after that.

Okay, but isn't that impractical for repeated builds?

If this is the official way to build llvm on a system with gcc-4.9 as
the host CXX, it's probably best to document it.