On Sun, May 07, 2017 at 05:37:47PM +0200, Mark Kettenis wrote:
> So this enables the code. It makes -femulated-tls the default,
> otherwise it will generated TLS relocations that we can't handle yet.
> It is possible to specify -fno-emulated-tls if you really want to
> generate those.
>
> A trivial example program that modifies the variable in one thread and
> prints it in another seems to work as expected.
>
> Note that the libcompiler_rt implementation unconditionally references
> some symbols in libpthread. So linking with -lpthread is mandatory.
> GCC 4.9 uses weak references, so -lpthread isn't needed there.
>
> Perhaps those functions should move to libc. Or we can make the
> libcompiler_rt code a bit smarter.
>
> Also note that the libcompiler_rt implementation offers the choice
> between posix_memalign() and normal malloc(). With this diff it uses
> malloc(). Perhaps otto@ has an opinion what's better here?
>
>
> Index: gnu/llvm/tools/clang/lib/Basic/Targets.cpp
> ===================================================================
> RCS file: /cvs/src/gnu/llvm/tools/clang/lib/Basic/Targets.cpp,v
> retrieving revision 1.7
> diff -u -p -r1.7 Targets.cpp
> --- gnu/llvm/tools/clang/lib/Basic/Targets.cpp 14 Mar 2017 08:11:34
> -0000 1.7
> +++ gnu/llvm/tools/clang/lib/Basic/Targets.cpp 7 May 2017 15:25:08
> -0000
> @@ -549,8 +549,6 @@ protected:
> public:
> OpenBSDTargetInfo(const llvm::Triple &Triple, const TargetOptions &Opts)
> : OSTargetInfo<Target>(Triple, Opts) {
> - this->TLSSupported = false;
> -
> switch (Triple.getArch()) {
> default:
> case llvm::Triple::x86:
> Index: gnu/llvm/tools/clang/lib/Driver/Tools.cpp
> ===================================================================
> RCS file: /cvs/src/gnu/llvm/tools/clang/lib/Driver/Tools.cpp,v
> retrieving revision 1.10
> diff -u -p -r1.10 Tools.cpp
> --- gnu/llvm/tools/clang/lib/Driver/Tools.cpp 24 Jan 2017 08:39:08 -0000
> 1.10
> +++ gnu/llvm/tools/clang/lib/Driver/Tools.cpp 7 May 2017 15:25:09 -0000
> @@ -5370,7 +5370,8 @@ void Clang::ConstructJob(Compilation &C,
> Args.AddLastArg(CmdArgs, options::OPT_fno_operator_names);
> // Emulated TLS is enabled by default on Android, and can be enabled
> manually
> // with -femulated-tls.
> - bool EmulatedTLSDefault = Triple.isAndroid() ||
> Triple.isWindowsCygwinEnvironment();
> + bool EmulatedTLSDefault = Triple.isAndroid() || Triple.isOSOpenBSD() ||
> + Triple.isWindowsCygwinEnvironment();
> if (Args.hasFlag(options::OPT_femulated_tls, options::OPT_fno_emulated_tls,
> EmulatedTLSDefault))
> CmdArgs.push_back("-femulated-tls");
> Index: lib/libcompiler_rt/Makefile
> ===================================================================
> RCS file: /cvs/src/lib/libcompiler_rt/Makefile,v
> retrieving revision 1.7
> diff -u -p -r1.7 Makefile
> --- lib/libcompiler_rt/Makefile 17 Apr 2017 15:53:21 -0000 1.7
> +++ lib/libcompiler_rt/Makefile 7 May 2017 15:25:09 -0000
> @@ -64,6 +64,7 @@ GEN_SRCS= absvdi2 \
> divti3 \
> divtf3 \
> divxc3 \
> + emutls \
> enable_execute_stack \
> eprintf \
> extendsfdf2 \
I think it can go in now. It allows more code in the ports tree to build,
and there's no way it can be fully tested until it's actually used, so having
it on by default it good. I see no drawback to this. Not that tls actually
doesn't work unless you have -std=c++11 at least.
