[SystemZ][z/OS] add code for hardware_concurrency()

[muiez]

Summary

Use the CSD_NUMBER_ONLINE_CPUS entry to return the hardware_concurrency for MVS.

[zibi]

LGTM

[ldionne]

libcxx/src/thread.cpp

110

There is no named API for this on z/OS?

[muiez]

libcxx/src/thread.cpp

110

Unfortunately, there isn't an API for this on z/OS.

[muiez]

Ping :) @ldionne

[muiez]

Move z/OS hardware_concurrency implementation to support/ibm directory.

[zibi]

libcxx/src/support/ibm/hardware_concurrency_zos.cpp

38

Why don't you rename this to hardware_concurrency_zos() , make it static function and call it from the original place in thread.cpp?

[muiez]

Refactor

[zibi]

LGTM

libcxx/include/thread

275

This could be non-member static function in ibm namespace, same as we did nonosleep()..

[philnik]

libcxx/include/thread

275

You have to __uglify the name.

[zibi]

libcxx/include/thread

275

Are you refering to hardware_concurrency_zos() or nonosleep()?
Sorry about the typo, it should be nanosleep().

Do you have any suggestion if you refer to hardware_concurrency_zos() ?

[ldionne]

I am extremely concerned by the need to introduce these sorts of arcane implementation details in the library for z/OS, especially since we don't have a clear idea of where these patches are going to stop.

libc++ is not a "low level system library". It's a library layered on top of C libraries and other standard (or mostly standard) APIs like POSIX and others. It's fine to have some amount of low level and platform specific stuff, but we're not going to implement from scratch APIs that should be provided by any decent OS.

libcxx/include/thread

274

This can't have inline visibility if it's in the built library.

275

And yeah it shouldn't be in the thread class either way.

libcxx/src/support/ibm/hardware_concurrency_zos.cpp

1

We don't put file names here anymore.

[muiez]

In D134216#3891898, @ldionne wrote:

I am extremely concerned by the need to introduce these sorts of arcane implementation details in the library for z/OS, especially since we don't have a clear idea of where these patches are going to stop.

libc++ is not a "low level system library". It's a library layered on top of C libraries and other standard (or mostly standard) APIs like POSIX and others. It's fine to have some amount of low level and platform specific stuff, but we're not going to implement from scratch APIs that should be provided by any decent OS.

Thanks for the feedback. I'll close this revision until we have an API for this on z/OS. It should be noted that it is in the works but not feasible in the short term, which is why we upstreamed this workaround in the meantime. To be more upstream friendly, we wanted to keep it in the support/ibm directory until we have an API.

libcxx/include/thread

#endif
     void join();
     void detach();
     _LIBCPP_INLINE_VISIBILITY
     id get_id() const _NOEXCEPT {return __libcpp_thread_get_id(&__t_);}
     _LIBCPP_INLINE_VISIBILITY
     native_handle_type native_handle() _NOEXCEPT {return __t_;}
 
     static unsigned hardware_concurrency() _NOEXCEPT;
+#ifdef __MVS__
+    _LIBCPP_INLINE_VISIBILITY
+    static unsigned hardware_concurrency_zos() _NOEXCEPT;
+#endif
 };
 
 #ifndef _LIBCPP_CXX03_LANG
 
 template <class _TSp, class _Fp, class ..._Args, size_t ..._Indices>
 inline _LIBCPP_INLINE_VISIBILITY
 void
 __thread_execute(tuple<_TSp, _Fp, _Args...>& __t, __tuple_indices<_Indices...>)

libcxx/src/CMakeLists.txt

list(APPEND LIBCXX_SOURCES
     support/solaris/wcsnrtombs.inc
     support/solaris/xlocale.cpp
     )
 elseif(ZOS)
   list(APPEND LIBCXX_SOURCES
     support/ibm/mbsnrtowcs.cpp
     support/ibm/wcsnrtombs.cpp
     support/ibm/xlocale_zos.cpp
+    support/ibm/hardware_concurrency_zos.cpp
     )
 endif()
 
 if (LIBCXX_ENABLE_FILESYSTEM)
   list(APPEND LIBCXX_SOURCES
     filesystem/filesystem_common.h
     filesystem/operations.cpp
     filesystem/directory_iterator.cpp

libcxx/src/support/ibm/hardware_concurrency_zos.cpp

+//===------------------------- hardware_concurrency_zos.cpp----------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+//===----------------------------------------------------------------------===//
+
+#include <__config>
+
+#ifndef _LIBCPP_HAS_NO_THREADS
+#include <thread>
+
+_LIBCPP_BEGIN_NAMESPACE_STD
+
+// The following offsets are from the MVS Data Areas Volumes 1 & 3:
+// Volume 1: https://www-01.ibm.com/servers/resourcelink/svc00100.nsf/pages/
+//           zosv2r3ga320935/$file/iead100_v2r3.pdf
+// Volume 3: https://www-01.ibm.com/servers/resourcelink/svc00100.nsf/pages/
+//           zosv2r3ga320938/$file/iead300_v2r3.pdf
+//
+// In Vol 1, CSD_NUMBER_ONLINE_CPUS is at decimal offset 212 in CSD Mapping
+// (also Vol 1), which is pointed to by CVTCSD.
+//
+// CVTCSD is at decimal offset 660 in CVT Mapping (Vol 1), which is pointed
+// to by FLCCVT field of the PSA data area.
+//
+// FLCCVT is at decimal offset 16 in PSA Mapping (Vol 3), and PSA is at
+// address 0.
+//
+// So, with the 32-bit pointer casts, CSD_NUMBER_ONLINE_CPUS is at
+// 0[16/4][660/4][212/4] which is 0[4][165][53].
+//
+// Note: z/OS is always upward compatible, so these offsets won't change.
+
+unsigned
+thread::hardware_concurrency_zos() noexcept {
+  return ((((int* __ptr32* __ptr32* __ptr32)0)[4])[165])[53];
+}
+
+_LIBCPP_END_NAMESPACE_STD
+
+#endif // !_LIBCPP_HAS_NO_THREADS

libcxx/src/thread.cpp

#if defined(_SC_NPROCESSORS_ONLN)
     // what is going on. Default to something safe.
     if (result < 0)
         return 0;
     return static_cast<unsigned>(result);
 #elif defined(_LIBCPP_WIN32API)
     SYSTEM_INFO info;
     GetSystemInfo(&info);
     return info.dwNumberOfProcessors;
+#elif defined(__MVS__)
+    return thread::hardware_concurrency_zos();
 #else  // defined(CTL_HW) && defined(HW_NCPU)
     // TODO: grovel through /proc or check cpuid on x86 and similar
     // instructions on other architectures.
 #   if defined(_LIBCPP_WARNING)
         _LIBCPP_WARNING("hardware_concurrency not yet implemented")
 #   else
 #       warning hardware_concurrency not yet implemented
 #   endif
     return 0;  // Means not computable [thread.thread.static]
 #endif // defined(CTL_HW) && defined(HW_NCPU)
 }
 
 namespace this_thread
 {
 
 void
 sleep_for(const chrono::nanoseconds& ns)
 {
     if (ns > chrono::nanoseconds::zero())
     {
         __libcpp_thread_sleep_for(ns);
     }
 }
 
 }  // this_thread
 
 __thread_specific_ptr<__thread_struct>&
 __thread_local_data()