Re: [llvm-dev] llvm-toolchain-3.8 on lower arm targets

On Wed, Oct 5, 2016 at 7:46 AM, Tim Northover via llvm-dev <llvm-dev@lists.llvm.org> wrote:

Hi Emilio,

On 4 October 2016 at 11:14, Emilio Pozuelo Monfort via llvm-dev
<llvm-dev@lists.llvm.org> wrote:
> In file included from /«PKGBUILDDIR»/lib/Support/ThreadPool.cpp:14:0:
> /«PKGBUILDDIR»/include/llvm/Support/ThreadPool.h: In member function
> 'std::shared_future<void> llvm::ThreadPool::async(Function&&, Args&& ...)':
> /«PKGBUILDDIR»/include/llvm/Support/ThreadPool.h:78:77: error: return type
> 'class std::shared_future<void>' is incomplete
> inline std::shared_future<VoidTy> async(Function &&F, Args &&... ArgList) {
> ^
>
> Any idea about this failure?
>
> For the Debian armel porters, we're switching to LLVM 3.8, so this failure
> (which happens on 3.8, 3.9 and llvm-toolchain-snapshot) is likely going to cause
> some package removals on armel as we try to get rid of older LLVM versions.
> Helping fixing this issue would be appreciated to prevent that.

This looks like the kind of failure you get when your host toolchain
doesn't support C++11 properly (specifically lock-free atomics in this
case). When I've seen it before GCC was defaulting to a CPU that's
too old to do atomics properly, and that configuration is very
unlikely to be supported by LLVM ever (any more).

This seems bogus.

C++11 allows atomic variables to be implemented using mutexes if the CPU doesn't support atomic operations on a given data type in some other way.

If you don't call atomic_is_lock_free(&var) then everything should work correctly, albeit perhaps more slowly than you might like.

It seems to me that atomic_is_lock_free() (or precomputed shortcuts such as ATOMIC_INT_LOCK_FREE) is there to enable you the possibility to use a different algorithm (if one is available), not to throw up your hands and say you don't support that architecture at all!

If it's the standard library going out of its way to check ATOMIC_INT_LOCK_FREE and then throwing up its hands and giving up then I'd say that's a bug. Simply taking out that check should produce working, correct code on anything that supports mutexes at all.