Implement std::pmr::synchronized_pool_resource

Message ID 20181108180229.GA31340@redhat.com
State New
Headers show
Series
  • Implement std::pmr::synchronized_pool_resource
Related show

Commit Message

Jonathan Wakely Nov. 8, 2018, 6:02 p.m.
Define the thread-safe pool resource, using a shared_mutex to allow
multiple threads to concurrently allocate from thread-specific pools.

Define new weak symbols for the pthread_rwlock_t functions, to avoid
making libstdc++.so depend on libpthread.so

	* config/abi/pre/gnu.ver: Add new symbols.
	* include/std/memory_resource [_GLIBCXX_HAS_GTHREADS]
	(synchronized_pool_resource): New class.
	* include/std/shared_mutex (__glibcxx_rwlock_rdlock)
	(__glibcxx_rwlock_tryrdlock, __glibcxx_rwlock_wrlock)
	(__glibcxx_rwlock_trywrlock, __glibcxx_rwlock_unlock)
	(__glibcxx_rwlock_destroy, __glibcxx_rwlock_init)
	(__glibcxx_rwlock_timedrdlock, __glibcxx_rwlock_timedwrlock): Define
	weak symbols for POSIX rwlock functions.
	(__shared_mutex_pthread): Use weak symbols.
	* src/c++17/memory_resource.cc [_GLIBCXX_HAS_GTHREADS]
	(synchronized_pool_resource::_TPools): New class.
	(destroy_TPools): New function for pthread_key_create destructor.
	(synchronized_pool_resource::synchronized_pool_resource)
	(synchronized_pool_resource::~synchronized_pool_resource)
	(synchronized_pool_resource::release)
	(synchronized_pool_resource::do_allocate)
	(synchronized_pool_resource::do_deallocate): Define public members.
	(synchronized_pool_resource::_M_thread_specific_pools)
	(synchronized_pool_resource::_M_alloc_tpools)
	(synchronized_pool_resource::_M_alloc_shared_tpools): Define private
	members.

The performance of this implementation is ... not great. But it's
better than simply adding a mutex to unsynchronized_pool_resource and
locking that around every allocation and deallocation.

I am not committing this yet, because I still need to test on non-GNU
targets, where the new weak wrappers around the pthread_rwlock_t
functions might not work properly.
commit adf32f3dd71a8312f212aadb7df2b6d32d8fe59b
Author: Jonathan Wakely <jwakely@redhat.com>
Date:   Sat Nov 3 23:13:26 2018 +0000

    Implement std::pmr::synchronized_pool_resource
    
    Define the thread-safe pool resource, using a shared_mutex to allow
    multiple threads to concurrently allocate from thread-specific pools.
    
    Define new weak symbols for the pthread_rwlock_t functions, to avoid
    making libstdc++.so depend on libpthread.so
    
            * config/abi/pre/gnu.ver: Add new symbols.
            * include/std/memory_resource [_GLIBCXX_HAS_GTHREADS]
            (synchronized_pool_resource): New class.
            * include/std/shared_mutex (__glibcxx_rwlock_rdlock)
            (__glibcxx_rwlock_tryrdlock, __glibcxx_rwlock_wrlock)
            (__glibcxx_rwlock_trywrlock, __glibcxx_rwlock_unlock)
            (__glibcxx_rwlock_destroy, __glibcxx_rwlock_init)
            (__glibcxx_rwlock_timedrdlock, __glibcxx_rwlock_timedwrlock): Define
            weak symbols for POSIX rwlock functions.
            (__shared_mutex_pthread): Use weak symbols.
            * src/c++17/memory_resource.cc [_GLIBCXX_HAS_GTHREADS]
            (synchronized_pool_resource::_TPools): New class.
            (destroy_TPools): New function for pthread_key_create destructor.
            (synchronized_pool_resource::synchronized_pool_resource)
            (synchronized_pool_resource::~synchronized_pool_resource)
            (synchronized_pool_resource::release)
            (synchronized_pool_resource::do_allocate)
            (synchronized_pool_resource::do_deallocate): Define public members.
            (synchronized_pool_resource::_M_thread_specific_pools)
            (synchronized_pool_resource::_M_alloc_tpools)
            (synchronized_pool_resource::_M_alloc_shared_tpools): Define private
            members.

Comments

Jonathan Wakely Nov. 17, 2018, 11:35 p.m. | #1
On 08/11/18 18:02 +0000, Jonathan Wakely wrote:
>Define the thread-safe pool resource, using a shared_mutex to allow
>multiple threads to concurrently allocate from thread-specific pools.
>
>Define new weak symbols for the pthread_rwlock_t functions, to avoid
>making libstdc++.so depend on libpthread.so
>
>	* config/abi/pre/gnu.ver: Add new symbols.
>	* include/std/memory_resource [_GLIBCXX_HAS_GTHREADS]
>	(synchronized_pool_resource): New class.
>	* include/std/shared_mutex (__glibcxx_rwlock_rdlock)
>	(__glibcxx_rwlock_tryrdlock, __glibcxx_rwlock_wrlock)
>	(__glibcxx_rwlock_trywrlock, __glibcxx_rwlock_unlock)
>	(__glibcxx_rwlock_destroy, __glibcxx_rwlock_init)
>	(__glibcxx_rwlock_timedrdlock, __glibcxx_rwlock_timedwrlock): Define
>	weak symbols for POSIX rwlock functions.
>	(__shared_mutex_pthread): Use weak symbols.
>	* src/c++17/memory_resource.cc [_GLIBCXX_HAS_GTHREADS]
>	(synchronized_pool_resource::_TPools): New class.
>	(destroy_TPools): New function for pthread_key_create destructor.
>	(synchronized_pool_resource::synchronized_pool_resource)
>	(synchronized_pool_resource::~synchronized_pool_resource)
>	(synchronized_pool_resource::release)
>	(synchronized_pool_resource::do_allocate)
>	(synchronized_pool_resource::do_deallocate): Define public members.
>	(synchronized_pool_resource::_M_thread_specific_pools)
>	(synchronized_pool_resource::_M_alloc_tpools)
>	(synchronized_pool_resource::_M_alloc_shared_tpools): Define private
>	members.
>
>The performance of this implementation is ... not great. But it's
>better than simply adding a mutex to unsynchronized_pool_resource and
>locking that around every allocation and deallocation.
>
>I am not committing this yet, because I still need to test on non-GNU
>targets, where the new weak wrappers around the pthread_rwlock_t
>functions might not work properly.

Here's the updated patch I'm committing, which includes more
conformance tests, and also performance tests.

Tested x86_64-linux, powerpc64le-linux, x86_64-freebsd11.0,
x86_64-w64-mingw32 and powerpc-aix7.2.0.0, committed to trunk.
commit 273ae309f8ee48d495054d0981da5d604bd36517
Author: Jonathan Wakely <jwakely@redhat.com>
Date:   Sat Nov 3 23:13:26 2018 +0000

    Implement std::pmr::synchronized_pool_resource
    
    Define the thread-safe pool resource, using a shared_mutex to allow
    multiple threads to concurrently allocate from thread-specific pools.
    
    Define new weak symbols for the pthread_rwlock_t functions, to avoid
    making libstdc++.so depend on libpthread.so
    
    When the necessary Gthread support is absent only define the
    feature-test macro to 1, rather than 201603. This is intended to imply
    incomplete support, because everything except synchronized_pool_resource
    works.
    
            Implement std::pmr::synchronized_pool_resource
            * config/abi/pre/gnu.ver: Add new symbols.
            * include/std/memory_resource [_GLIBCXX_HAS_GTHREADS]
            (__cpp_lib_memory_resource): Define to expected value, 201603.
            (synchronized_pool_resource): New class.
            [!_GLIBCXX_HAS_GTHREADS] (__cpp_lib_memory_resource): Define to 1.
            * include/std/shared_mutex (__glibcxx_rwlock_rdlock)
            (__glibcxx_rwlock_tryrdlock, __glibcxx_rwlock_wrlock)
            (__glibcxx_rwlock_trywrlock, __glibcxx_rwlock_unlock)
            (__glibcxx_rwlock_destroy, __glibcxx_rwlock_init)
            (__glibcxx_rwlock_timedrdlock, __glibcxx_rwlock_timedwrlock): Define
            weak symbols for POSIX rwlock functions.
            (__shared_mutex_pthread): Use weak symbols.
            * include/std/version (__cpp_lib_memory_resource): Define.
            * src/c++17/memory_resource.cc [_GLIBCXX_HAS_GTHREADS]
            (synchronized_pool_resource::_TPools): New class.
            (destroy_TPools): New function for pthread_key_create destructor.
            (synchronized_pool_resource::synchronized_pool_resource)
            (synchronized_pool_resource::~synchronized_pool_resource)
            (synchronized_pool_resource::release)
            (synchronized_pool_resource::do_allocate)
            (synchronized_pool_resource::do_deallocate): Define public members.
            (synchronized_pool_resource::_M_thread_specific_pools)
            (synchronized_pool_resource::_M_alloc_tpools)
            (synchronized_pool_resource::_M_alloc_shared_tpools): Define private
            members.
            * testsuite/20_util/synchronized_pool_resource/allocate.cc: New test.
            * testsuite/20_util/synchronized_pool_resource/cons.cc: New test.
            * testsuite/20_util/synchronized_pool_resource/is_equal.cc: New test.
            * testsuite/20_util/synchronized_pool_resource/multithreaded.cc: New
            test.
            * testsuite/20_util/synchronized_pool_resource/release.cc: New test.
            * testsuite/performance/20_util/memory_resource/pools.cc: Add
            multithreaded tests using pmr::synchronized_pool_resource.

diff --git a/libstdc++-v3/config/abi/pre/gnu.ver b/libstdc++-v3/config/abi/pre/gnu.ver
index 9d66f908e1a..e4bda4f6f1e 100644
--- a/libstdc++-v3/config/abi/pre/gnu.ver
+++ b/libstdc++-v3/config/abi/pre/gnu.ver
@@ -2039,13 +2039,6 @@ GLIBCXX_3.4.26 {
     _ZNSt7__cxx1118basic_stringstreamI[cw]St11char_traitsI[cw]ESaI[cw]EEC[12]Ev;
     _ZNSt7__cxx1119basic_[io]stringstreamI[cw]St11char_traitsI[cw]ESaI[cw]EEC[12]Ev;
 
-    _ZNSt3pmr19new_delete_resourceEv;
-    _ZNSt3pmr20null_memory_resourceEv;
-    _ZNSt3pmr20get_default_resourceEv;
-    _ZNSt3pmr20set_default_resourceEPNS_15memory_resourceE;
-    _ZNSt3pmr25monotonic_buffer_resource13_M_new_bufferE[jmy][jmy];
-    _ZNSt3pmr25monotonic_buffer_resource18_M_release_buffersEv;
-
     # std::__throw_ios_failure(const char*, int);
     _ZSt19__throw_ios_failurePKci;
 
@@ -2057,6 +2050,18 @@ GLIBCXX_3.4.26 {
     _ZN11__gnu_debug25_Safe_local_iterator_base16_M_attach_singleEPNS_19_Safe_sequence_baseEb;
 
     # <memory_resource> members
+    _ZNSt3pmr19new_delete_resourceEv;
+    _ZNSt3pmr20null_memory_resourceEv;
+    _ZNSt3pmr20get_default_resourceEv;
+    _ZNSt3pmr20set_default_resourceEPNS_15memory_resourceE;
+    _ZNSt3pmr25monotonic_buffer_resource13_M_new_bufferE[jmy][jmy];
+    _ZNSt3pmr25monotonic_buffer_resource18_M_release_buffersEv;
+    _ZTINSt3pmr26synchronized_pool_resourceE;
+    _ZNSt3pmr26synchronized_pool_resourceC[12]ERKNS_12pool_optionsEPNS_15memory_resourceE;
+    _ZNSt3pmr26synchronized_pool_resourceD[12]Ev;
+    _ZNSt3pmr26synchronized_pool_resource7releaseEv;
+    _ZNSt3pmr26synchronized_pool_resource11do_allocateE[jmy][jmy];
+    _ZNSt3pmr26synchronized_pool_resource13do_deallocateEPv[jmy][jmy];
     _ZTINSt3pmr28unsynchronized_pool_resourceE;
     _ZNSt3pmr28unsynchronized_pool_resourceC[12]ERKNS_12pool_optionsEPNS_15memory_resourceE;
     _ZNSt3pmr28unsynchronized_pool_resourceD[12]Ev;
diff --git a/libstdc++-v3/include/std/memory_resource b/libstdc++-v3/include/std/memory_resource
index 40486af82fe..87ad25d60f3 100644
--- a/libstdc++-v3/include/std/memory_resource
+++ b/libstdc++-v3/include/std/memory_resource
@@ -37,6 +37,7 @@
 #include <utility>			// pair, index_sequence
 #include <vector>			// vector
 #include <cstddef>			// size_t, max_align_t
+#include <shared_mutex>			// shared_mutex
 #include <debug/assertions.h>
 
 namespace std _GLIBCXX_VISIBILITY(default)
@@ -44,7 +45,13 @@ namespace std _GLIBCXX_VISIBILITY(default)
 _GLIBCXX_BEGIN_NAMESPACE_VERSION
 namespace pmr
 {
-// #define __cpp_lib_memory_resource 201603
+#ifdef _GLIBCXX_HAS_GTHREADS
+  // Header and all contents are present.
+# define __cpp_lib_memory_resource 201603
+#else
+  // The pmr::synchronized_pool_resource type is missing.
+# define __cpp_lib_memory_resource 1
+#endif
 
   class memory_resource;
 
@@ -60,7 +67,9 @@ namespace pmr
 
   // Pool resource classes
   struct pool_options;
+#ifdef _GLIBCXX_HAS_GTHREADS
   class synchronized_pool_resource;
+#endif
   class unsynchronized_pool_resource;
   class monotonic_buffer_resource;
 
@@ -338,7 +347,72 @@ namespace pmr
     const int _M_npools;
   };
 
-  // TODO class synchronized_pool_resource
+#ifdef _GLIBCXX_HAS_GTHREADS
+  /// A thread-safe memory resource that manages pools of fixed-size blocks.
+  class synchronized_pool_resource : public memory_resource
+  {
+  public:
+    synchronized_pool_resource(const pool_options& __opts,
+				 memory_resource* __upstream)
+    __attribute__((__nonnull__));
+
+    synchronized_pool_resource()
+    : synchronized_pool_resource(pool_options(), get_default_resource())
+    { }
+
+    explicit
+    synchronized_pool_resource(memory_resource* __upstream)
+    __attribute__((__nonnull__))
+    : synchronized_pool_resource(pool_options(), __upstream)
+    { }
+
+    explicit
+    synchronized_pool_resource(const pool_options& __opts)
+    : synchronized_pool_resource(__opts, get_default_resource()) { }
+
+    synchronized_pool_resource(const synchronized_pool_resource&) = delete;
+
+    virtual ~synchronized_pool_resource();
+
+    synchronized_pool_resource&
+    operator=(const synchronized_pool_resource&) = delete;
+
+    void release();
+
+    memory_resource*
+    upstream_resource() const noexcept
+    __attribute__((__returns_nonnull__))
+    { return _M_impl.resource(); }
+
+    pool_options options() const noexcept { return _M_impl._M_opts; }
+
+  protected:
+    void*
+    do_allocate(size_t __bytes, size_t __alignment) override;
+
+    void
+    do_deallocate(void* __p, size_t __bytes, size_t __alignment) override;
+
+    bool
+    do_is_equal(const memory_resource& __other) const noexcept override
+    { return this == &__other; }
+
+  public:
+    // Thread-specific pools (only public for access by implementation details)
+    struct _TPools;
+
+  private:
+    _TPools* _M_alloc_tpools(lock_guard<shared_mutex>&);
+    _TPools* _M_alloc_shared_tpools(lock_guard<shared_mutex>&);
+    auto _M_thread_specific_pools() noexcept;
+
+    __pool_resource _M_impl;
+    __gthread_key_t _M_key;
+    // Linked list of thread-specific pools. All threads share _M_tpools[0].
+    _TPools* _M_tpools = nullptr;
+    mutable shared_mutex _M_mx;
+  };
+#endif
 
   /// A non-thread-safe memory resource that manages pools of fixed-size blocks.
   class unsynchronized_pool_resource : public memory_resource
diff --git a/libstdc++-v3/include/std/shared_mutex b/libstdc++-v3/include/std/shared_mutex
index dce97f48a3f..8aa6e9f0d4f 100644
--- a/libstdc++-v3/include/std/shared_mutex
+++ b/libstdc++-v3/include/std/shared_mutex
@@ -57,6 +57,90 @@ _GLIBCXX_BEGIN_NAMESPACE_VERSION
   class shared_timed_mutex;
 
 #if _GLIBCXX_USE_PTHREAD_RWLOCK_T
+#ifdef __gthrw
+#define _GLIBCXX_GTHRW(name) \
+  __gthrw(pthread_ ## name); \
+  static inline int \
+  __glibcxx_ ## name (pthread_rwlock_t *__rwlock) \
+  { \
+    if (__gthread_active_p ()) \
+      return __gthrw_(pthread_ ## name) (__rwlock); \
+    else \
+      return 0; \
+  }
+  _GLIBCXX_GTHRW(rwlock_rdlock)
+  _GLIBCXX_GTHRW(rwlock_tryrdlock)
+  _GLIBCXX_GTHRW(rwlock_wrlock)
+  _GLIBCXX_GTHRW(rwlock_trywrlock)
+  _GLIBCXX_GTHRW(rwlock_unlock)
+# ifndef PTHREAD_RWLOCK_INITIALIZER
+  _GLIBCXX_GTHRW(rwlock_destroy)
+  __gthrw(pthread_rwlock_init);
+  static inline int
+  __glibcxx_rwlock_init (pthread_rwlock_t *__rwlock)
+  {
+    if (__gthread_active_p ())
+      return __gthrw_(pthread_rwlock_init) (__rwlock, NULL);
+    else
+      return 0;
+  }
+# endif
+# if _GTHREAD_USE_MUTEX_TIMEDLOCK
+   __gthrw(pthread_rwlock_timedrdlock);
+  static inline int
+  __glibcxx_rwlock_timedrdlock (pthread_rwlock_t *__rwlock,
+				const timespec *__ts)
+  {
+    if (__gthread_active_p ())
+      return __gthrw_(pthread_rwlock_timedrdlock) (__rwlock, __ts);
+    else
+      return 0;
+  }
+   __gthrw(pthread_rwlock_timedwrlock);
+  static inline int
+  __glibcxx_rwlock_timedwrlock (pthread_rwlock_t *__rwlock,
+				const timespec *__ts)
+  {
+    if (__gthread_active_p ())
+      return __gthrw_(pthread_rwlock_timedwrlock) (__rwlock, __ts);
+    else
+      return 0;
+  }
+# endif
+#else
+  static inline int
+  __glibcxx_rwlock_rdlock (pthread_rwlock_t *__rwlock)
+  { return pthread_rwlock_rdlock (__rwlock); }
+  static inline int
+  __glibcxx_rwlock_tryrdlock (pthread_rwlock_t *__rwlock)
+  { return pthread_rwlock_tryrdlock (__rwlock); }
+  static inline int
+  __glibcxx_rwlock_wrlock (pthread_rwlock_t *__rwlock)
+  { return pthread_rwlock_wrlock (__rwlock); }
+  static inline int
+  __glibcxx_rwlock_trywrlock (pthread_rwlock_t *__rwlock)
+  { return pthread_rwlock_trywrlock (__rwlock); }
+  static inline int
+  __glibcxx_rwlock_unlock (pthread_rwlock_t *__rwlock)
+  { return pthread_rwlock_unlock (__rwlock); }
+  static inline int
+  __glibcxx_rwlock_destroy(pthread_rwlock_t *__rwlock)
+  { return pthread_rwlock_destroy (__rwlock); }
+  static inline int
+  __glibcxx_rwlock_init(pthread_rwlock_t *__rwlock)
+  { return pthread_rwlock_init (__rwlock, NULL); }
+# if _GTHREAD_USE_MUTEX_TIMEDLOCK
+  static inline int
+  __glibcxx_rwlock_timedrdlock (pthread_rwlock_t *__rwlock,
+				const timespec *__ts)
+  { return pthread_rwlock_timedrdlock (__rwlock, __ts); }
+  static inline int
+  __glibcxx_rwlock_timedwrlock (pthread_rwlock_t *__rwlock,
+				const timespec *__ts)
+  { return pthread_rwlock_timedwrlock (__rwlock, __ts); }
+# endif
+#endif
+
   /// A shared mutex type implemented using pthread_rwlock_t.
   class __shared_mutex_pthread
   {
@@ -74,7 +158,7 @@ _GLIBCXX_BEGIN_NAMESPACE_VERSION
   public:
     __shared_mutex_pthread()
     {
-      int __ret = pthread_rwlock_init(&_M_rwlock, NULL);
+      int __ret = __glibcxx_rwlock_init(&_M_rwlock, NULL);
       if (__ret == ENOMEM)
 	__throw_bad_alloc();
       else if (__ret == EAGAIN)
@@ -87,7 +171,7 @@ _GLIBCXX_BEGIN_NAMESPACE_VERSION
 
     ~__shared_mutex_pthread()
     {
-      int __ret __attribute((__unused__)) = pthread_rwlock_destroy(&_M_rwlock);
+      int __ret __attribute((__unused__)) = __glibcxx_rwlock_destroy(&_M_rwlock);
       // Errors not handled: EBUSY, EINVAL
       __glibcxx_assert(__ret == 0);
     }
@@ -99,7 +183,7 @@ _GLIBCXX_BEGIN_NAMESPACE_VERSION
     void
     lock()
     {
-      int __ret = pthread_rwlock_wrlock(&_M_rwlock);
+      int __ret = __glibcxx_rwlock_wrlock(&_M_rwlock);
       if (__ret == EDEADLK)
 	__throw_system_error(int(errc::resource_deadlock_would_occur));
       // Errors not handled: EINVAL
@@ -109,7 +193,7 @@ _GLIBCXX_BEGIN_NAMESPACE_VERSION
     bool
     try_lock()
     {
-      int __ret = pthread_rwlock_trywrlock(&_M_rwlock);
+      int __ret = __glibcxx_rwlock_trywrlock(&_M_rwlock);
       if (__ret == EBUSY) return false;
       // Errors not handled: EINVAL
       __glibcxx_assert(__ret == 0);
@@ -119,7 +203,7 @@ _GLIBCXX_BEGIN_NAMESPACE_VERSION
     void
     unlock()
     {
-      int __ret __attribute((__unused__)) = pthread_rwlock_unlock(&_M_rwlock);
+      int __ret __attribute((__unused__)) = __glibcxx_rwlock_unlock(&_M_rwlock);
       // Errors not handled: EPERM, EBUSY, EINVAL
       __glibcxx_assert(__ret == 0);
     }
@@ -135,7 +219,7 @@ _GLIBCXX_BEGIN_NAMESPACE_VERSION
       // is okay based on the current specification of forward progress
       // guarantees by the standard.
       do
-	__ret = pthread_rwlock_rdlock(&_M_rwlock);
+	__ret = __glibcxx_rwlock_rdlock(&_M_rwlock);
       while (__ret == EAGAIN);
       if (__ret == EDEADLK)
 	__throw_system_error(int(errc::resource_deadlock_would_occur));
@@ -146,7 +230,7 @@ _GLIBCXX_BEGIN_NAMESPACE_VERSION
     bool
     try_lock_shared()
     {
-      int __ret = pthread_rwlock_tryrdlock(&_M_rwlock);
+      int __ret = __glibcxx_rwlock_tryrdlock(&_M_rwlock);
       // If the maximum number of read locks has been exceeded, we just fail
       // to acquire the lock.  Unlike for lock(), we are not allowed to throw
       // an exception.
@@ -413,7 +497,7 @@ _GLIBCXX_BEGIN_NAMESPACE_VERSION
 	    static_cast<long>(__ns.count())
 	  };
 
-	int __ret = pthread_rwlock_timedwrlock(&_M_rwlock, &__ts);
+	int __ret = __glibcxx_rwlock_timedwrlock(&_M_rwlock, &__ts);
 	// On self-deadlock, we just fail to acquire the lock.  Technically,
 	// the program violated the precondition.
 	if (__ret == ETIMEDOUT || __ret == EDEADLK)
@@ -466,7 +550,7 @@ _GLIBCXX_BEGIN_NAMESPACE_VERSION
 	// mistaken for a spurious failure, which might help users realise
 	// there is a deadlock.
 	do
-	  __ret = pthread_rwlock_timedrdlock(&_M_rwlock, &__ts);
+	  __ret = __glibcxx_rwlock_timedrdlock(&_M_rwlock, &__ts);
 	while (__ret == EAGAIN || __ret == EDEADLK);
 	if (__ret == ETIMEDOUT)
 	  return false;
diff --git a/libstdc++-v3/include/std/version b/libstdc++-v3/include/std/version
index 0e054c47bf8..95950d82c31 100644
--- a/libstdc++-v3/include/std/version
+++ b/libstdc++-v3/include/std/version
@@ -118,6 +118,11 @@
 #define __cpp_lib_map_insertion 201411
 #define __cpp_lib_map_try_emplace 201411
 #define __cpp_lib_math_special_functions 201603L
+#ifdef _GLIBCXX_HAS_GTHREADS
+# define __cpp_lib_memory_resource 201603
+#else
+# define __cpp_lib_memory_resource 1
+#endif
 #define __cpp_lib_node_extract 201606
 #define __cpp_lib_nonmember_container_access 201411
 #define __cpp_lib_not_fn 201603
diff --git a/libstdc++-v3/src/c++17/memory_resource.cc b/libstdc++-v3/src/c++17/memory_resource.cc
index 79c1665146d..a3766311725 100644
--- a/libstdc++-v3/src/c++17/memory_resource.cc
+++ b/libstdc++-v3/src/c++17/memory_resource.cc
@@ -896,6 +896,11 @@ namespace pmr
     return n + 1;
   }
 
+#ifdef _GLIBCXX_HAS_GTHREADS
+  using shared_lock = std::shared_lock<shared_mutex>;
+  using exclusive_lock = lock_guard<shared_mutex>;
+#endif
+
   } // namespace
 
   __pool_resource::
@@ -987,6 +992,292 @@ namespace pmr
     return p;
   }
 
+#ifdef _GLIBCXX_HAS_GTHREADS
+  // synchronized_pool_resource members.
+
+  /* Notes on implementation and thread safety:
+   *
+   * Each synchronized_pool_resource manages an linked list of N+1 _TPools
+   * objects, where N is the number of threads using the pool resource.
+   * Each _TPools object has its own set of pools, with their own chunks.
+   * The first element of the list, _M_tpools[0], can be used by any thread.
+   * The rest of the list contains a _TPools object for each thread,
+   * accessed via the thread-specific key _M_key (and referred to for
+   * exposition as _M_tpools[_M_key]).
+   * The first element, _M_tpools[0], contains "orphaned chunks" which were
+   * allocated by a thread which has since exited, and so there is no
+   * _M_tpools[_M_key] for that thread.
+   * A thread can access its own thread-specific set of pools via _M_key
+   * while holding a shared lock on _M_mx. Accessing _M_impl._M_unpooled
+   * or _M_tpools[0] or any other thread's _M_tpools[_M_key] requires an
+   * exclusive lock.
+   * The upstream_resource() pointer can be obtained without a lock, but
+   * any dereference of that pointer requires an exclusive lock.
+   * The _M_impl._M_opts and _M_impl._M_npools members are immutable,
+   * and can safely be accessed concurrently.
+   */
+
+  extern "C" {
+    static void destroy_TPools(void*);
+  }
+
+  struct synchronized_pool_resource::_TPools
+  {
+    // Exclusive lock must be held in the thread where this constructor runs.
+    explicit
+    _TPools(synchronized_pool_resource& owner, exclusive_lock&)
+    : owner(owner), pools(owner._M_impl._M_alloc_pools())
+    {
+      // __builtin_printf("%p constructing\n", this);
+      __glibcxx_assert(pools);
+    }
+
+    // Exclusive lock must be held in the thread where this destructor runs.
+    ~_TPools()
+    {
+      __glibcxx_assert(pools);
+      if (pools)
+	{
+	  memory_resource* r = owner.upstream_resource();
+	  for (int i = 0; i < owner._M_impl._M_npools; ++i)
+	    pools[i].release(r);
+	  std::destroy_n(pools, owner._M_impl._M_npools);
+	  polymorphic_allocator<__pool_resource::_Pool> a(r);
+	  a.deallocate(pools, owner._M_impl._M_npools);
+	}
+      if (prev)
+	prev->next = next;
+      if (next)
+	next->prev = prev;
+    }
+
+    // Exclusive lock must be held in the thread where this function runs.
+    void move_nonempty_chunks()
+    {
+      __glibcxx_assert(pools);
+      if (!pools)
+	return;
+      memory_resource* r = owner.upstream_resource();
+      // move all non-empty chunks to the shared _TPools
+      for (int i = 0; i < owner._M_impl._M_npools; ++i)
+	for (auto& c : pools[i]._M_chunks)
+	  if (!c.empty())
+	    owner._M_tpools->pools[i]._M_chunks.insert(std::move(c), r);
+    }
+
+    synchronized_pool_resource& owner;
+    __pool_resource::_Pool* pools = nullptr;
+    _TPools* prev = nullptr;
+    _TPools* next = nullptr;
+
+    static void destroy(_TPools* p)
+    {
+      exclusive_lock l(p->owner._M_mx);
+      // __glibcxx_assert(p != p->owner._M_tpools);
+      p->move_nonempty_chunks();
+      polymorphic_allocator<_TPools> a(p->owner.upstream_resource());
+      p->~_TPools();
+      a.deallocate(p, 1);
+    }
+  };
+
+  // Called when a thread exits
+  extern "C" {
+    static void destroy_TPools(void* p)
+    {
+      using _TPools = synchronized_pool_resource::_TPools;
+      _TPools::destroy(static_cast<_TPools*>(p));
+    }
+  }
+
+  // Constructor
+  synchronized_pool_resource::
+  synchronized_pool_resource(const pool_options& opts,
+			     memory_resource* upstream)
+  : _M_impl(opts, upstream)
+  {
+    if (int err = __gthread_key_create(&_M_key, destroy_TPools))
+      __throw_system_error(err);
+    exclusive_lock l(_M_mx);
+    _M_tpools = _M_alloc_shared_tpools(l);
+  }
+
+  // Destructor
+  synchronized_pool_resource::~synchronized_pool_resource()
+  {
+    release();
+    __gthread_key_delete(_M_key); // does not run destroy_TPools
+  }
+
+  void
+  synchronized_pool_resource::release()
+  {
+    exclusive_lock l(_M_mx);
+    if (_M_tpools)
+      {
+	__gthread_key_delete(_M_key); // does not run destroy_TPools
+	__gthread_key_create(&_M_key, destroy_TPools);
+	polymorphic_allocator<_TPools> a(upstream_resource());
+	// destroy+deallocate each _TPools
+	do
+	  {
+	    _TPools* p = _M_tpools;
+	    _M_tpools = _M_tpools->next;
+	    p->~_TPools();
+	    a.deallocate(p, 1);
+	  }
+	while (_M_tpools);
+      }
+    // release unpooled memory
+    _M_impl.release();
+  }
+
+  // Caller must hold shared or exclusive lock to ensure the pointer
+  // isn't invalidated before it can be used.
+  auto
+  synchronized_pool_resource::_M_thread_specific_pools() noexcept
+  {
+    __pool_resource::_Pool* pools = nullptr;
+    if (auto tp = static_cast<_TPools*>(__gthread_getspecific(_M_key)))
+      {
+      pools = tp->pools;
+      __glibcxx_assert(tp->pools);
+      }
+    return pools;
+  }
+
+  // Override for memory_resource::do_allocate
+  void*
+  synchronized_pool_resource::
+  do_allocate(size_t bytes, size_t alignment)
+  {
+    const auto block_size = std::max(bytes, alignment);
+    if (block_size <= _M_impl._M_opts.largest_required_pool_block)
+      {
+	const ptrdiff_t index = pool_index(block_size, _M_impl._M_npools);
+	memory_resource* const r = upstream_resource();
+	const pool_options opts = _M_impl._M_opts;
+	{
+	  // Try to allocate from the thread-specific pool
+	  shared_lock l(_M_mx);
+	  if (auto pools = _M_thread_specific_pools()) // [[likely]]
+	    {
+	      // Need exclusive lock to replenish so use try_allocate:
+	      if (void* p = pools[index].try_allocate())
+		return p;
+	      // Need to take exclusive lock and replenish pool.
+	    }
+	  // Need to allocate or replenish thread-specific pools using
+	  // upstream resource, so need to hold exclusive lock.
+	}
+	// N.B. Another thread could call release() now lock is not held.
+	exclusive_lock excl(_M_mx);
+	if (!_M_tpools) // [[unlikely]]
+	  _M_tpools = _M_alloc_shared_tpools(excl);
+	auto pools = _M_thread_specific_pools();
+	if (!pools)
+	  pools = _M_alloc_tpools(excl)->pools;
+	return pools[index].allocate(r, opts);
+      }
+    exclusive_lock l(_M_mx);
+    return _M_impl.allocate(bytes, alignment); // unpooled allocation
+  }
+
+  // Override for memory_resource::do_deallocate
+  void
+  synchronized_pool_resource::
+  do_deallocate(void* p, size_t bytes, size_t alignment)
+  {
+    size_t block_size = std::max(bytes, alignment);
+    if (block_size <= _M_impl._M_opts.largest_required_pool_block)
+      {
+	const ptrdiff_t index = pool_index(block_size, _M_impl._M_npools);
+	__glibcxx_assert(index != -1);
+	{
+	  shared_lock l(_M_mx);
+	  auto pools = _M_thread_specific_pools();
+	  if (pools)
+	    {
+	      // No need to lock here, no other thread is accessing this pool.
+	      if (pools[index].deallocate(upstream_resource(), p))
+		return;
+	    }
+	  // Block might have come from a different thread's pool,
+	  // take exclusive lock and check every pool.
+	}
+	// TODO store {p, bytes, alignment} somewhere and defer returning
+	// the block to the correct thread-specific pool until we next
+	// take the exclusive lock.
+	exclusive_lock excl(_M_mx);
+	for (_TPools* t = _M_tpools; t != nullptr; t = t->next)
+	  {
+	    if (t->pools) // [[likely]]
+	      {
+		if (t->pools[index].deallocate(upstream_resource(), p))
+		  return;
+	      }
+	  }
+      }
+    exclusive_lock l(_M_mx);
+    _M_impl.deallocate(p, bytes, alignment);
+  }
+
+  // Allocate a thread-specific _TPools object and add it to the linked list.
+  auto
+  synchronized_pool_resource::_M_alloc_tpools(exclusive_lock& l)
+  -> _TPools*
+  {
+    __glibcxx_assert(_M_tpools != nullptr);
+    // dump_list(_M_tpools);
+    polymorphic_allocator<_TPools> a(upstream_resource());
+    _TPools* p = a.allocate(1);
+    bool constructed = false;
+    __try
+      {
+	a.construct(p, *this, l);
+	constructed = true;
+	// __glibcxx_assert(__gthread_getspecific(_M_key) == nullptr);
+	if (int err = __gthread_setspecific(_M_key, p))
+	  __throw_system_error(err);
+      }
+    __catch(...)
+      {
+	if (constructed)
+	  a.destroy(p);
+	a.deallocate(p, 1);
+	__throw_exception_again;
+      }
+    p->prev = _M_tpools;
+    p->next = _M_tpools->next;
+    _M_tpools->next = p;
+    if (p->next)
+      p->next->prev = p;
+    return p;
+  }
+
+  // Allocate the shared _TPools object, _M_tpools[0]
+  auto
+  synchronized_pool_resource::_M_alloc_shared_tpools(exclusive_lock& l)
+  -> _TPools*
+  {
+    __glibcxx_assert(_M_tpools == nullptr);
+    polymorphic_allocator<_TPools> a(upstream_resource());
+    _TPools* p = a.allocate(1);
+    __try
+      {
+	a.construct(p, *this, l);
+      }
+    __catch(...)
+      {
+	a.deallocate(p, 1);
+	__throw_exception_again;
+      }
+    // __glibcxx_assert(p->next == nullptr);
+    // __glibcxx_assert(p->prev == nullptr);
+    return p;
+  }
+#endif // _GLIBCXX_HAS_GTHREADS
+
   // unsynchronized_pool_resource member functions
 
   // Constructor
diff --git a/libstdc++-v3/testsuite/20_util/synchronized_pool_resource/allocate.cc b/libstdc++-v3/testsuite/20_util/synchronized_pool_resource/allocate.cc
new file mode 100644
index 00000000000..e2a0390e377
--- /dev/null
+++ b/libstdc++-v3/testsuite/20_util/synchronized_pool_resource/allocate.cc
@@ -0,0 +1,159 @@
+// Copyright (C) 2018 Free Software Foundation, Inc.
+//
+// This file is part of the GNU ISO C++ Library.  This library is free
+// software; you can redistribute it and/or modify it under the
+// terms of the GNU General Public License as published by the
+// Free Software Foundation; either version 3, or (at your option)
+// any later version.
+
+// This library is distributed in the hope that it will be useful,
+// but WITHOUT ANY WARRANTY; without even the implied warranty of
+// MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+// GNU General Public License for more details.
+
+// You should have received a copy of the GNU General Public License along
+// with this library; see the file COPYING3.  If not see
+// <http://www.gnu.org/licenses/>.
+
+// { dg-do run }
+// { dg-options "-std=gnu++17 -pthread" }
+// { dg-require-effective-target c++17 }
+// { dg-require-effective-target pthread }
+// { dg-require-gthreads "" }
+
+#include <memory_resource>
+#include <cstring>
+#include <testsuite_allocator.h>
+#include <testsuite_hooks.h>
+
+void
+test01()
+{
+  __gnu_test::memory_resource test_mr;
+  {
+    std::pmr::synchronized_pool_resource r(&test_mr);
+    void* p1 = r.allocate(1, 1);
+    VERIFY( p1 != nullptr );
+    auto n = test_mr.number_of_active_allocations();
+    VERIFY( n > 0 );
+    // Ensure memory region can be written to (without corrupting heap!)
+    std::memset(p1, 0xff, 1);
+    void* p2 = r.allocate(1, 1);
+    VERIFY( p2 != nullptr );
+    VERIFY( p2 != p1 );
+    VERIFY( test_mr.number_of_active_allocations() == n );
+    std::memset(p1, 0xff, 1);
+    r.deallocate(p1, 1, 1);
+    // Returning single blocks to the pool doesn't return them upstream:
+    VERIFY( test_mr.number_of_active_allocations() == n );
+    r.deallocate(p2, 1, 1);
+    VERIFY( test_mr.number_of_active_allocations() == n );
+  }
+  VERIFY( test_mr.number_of_active_allocations() == 0 );
+}
+
+void
+test02()
+{
+  struct nullable_memory_resource : public std::pmr::memory_resource
+  {
+    void*
+    do_allocate(std::size_t bytes, std::size_t alignment) override
+    { return upstream->allocate(bytes, alignment); }
+
+    void
+    do_deallocate(void* p, std::size_t bytes, std::size_t alignment) override
+    { upstream->deallocate(p, bytes, alignment); }
+
+    bool
+    do_is_equal(const memory_resource& r) const noexcept override
+    { return &r == this; }
+
+    std::pmr::memory_resource* upstream = std::pmr::get_default_resource();
+  };
+
+  nullable_memory_resource test_mr;
+  std::pmr::synchronized_pool_resource r(&test_mr);
+  void* p1 = r.allocate(8, 1);
+  VERIFY( p1 != nullptr );
+  std::memset(p1, 0xff, 8);
+  test_mr.upstream = nullptr;
+  void* p2 = r.allocate(8, 1); //should not need to replenish
+  VERIFY( p2 != nullptr );
+  VERIFY( p2 != p1 );
+  std::memset(p1, 0xff, 8);
+  r.deallocate(p1, 8, 1); // should not use upstream
+  r.deallocate(p2, 8, 1); // should not use upstream
+
+  // Destructor will return memory upstream, so restore the upstream resource:
+  test_mr.upstream = std::pmr::get_default_resource();
+}
+
+void
+test03()
+{
+  __gnu_test::memory_resource test_mr;
+  {
+    std::pmr::synchronized_pool_resource r({10, 16}, &test_mr);
+    std::size_t largest_pool = r.options().largest_required_pool_block;
+    void* p1 = r.allocate(2 * largest_pool);
+    VERIFY( p1 != nullptr );
+    const std::size_t n = test_mr.number_of_active_allocations();
+    // Allocation of pools + allocation of pmr::vector + oversize allocation:
+    VERIFY( n >= 1 );
+    std::memset(p1, 0xff, 2 * largest_pool);
+    void* p2 = r.allocate(3 * largest_pool);
+    VERIFY( p2 != nullptr );
+    VERIFY( p2 != p1 );
+    VERIFY( test_mr.number_of_active_allocations() == n + 1 );
+    std::memset(p2, 0xff, 3 * largest_pool);
+    r.deallocate(p1, 2 * largest_pool);
+    VERIFY( test_mr.number_of_active_allocations() ==  n );
+    r.deallocate(p2, 3 * largest_pool);
+    VERIFY( test_mr.number_of_active_allocations() == n - 1 );
+  }
+  VERIFY( test_mr.number_of_active_allocations() == 0 );
+  {
+    std::pmr::synchronized_pool_resource r({16, 16}, &test_mr);
+    (void) r.allocate(2);
+    (void) r.allocate(8);
+    (void) r.allocate(16);
+    (void) r.allocate(2);
+    (void) r.allocate(8);
+    (void) r.allocate(16);
+    (void) r.allocate(2 * r.options().largest_required_pool_block);
+    VERIFY( test_mr.number_of_active_allocations() != 0 );
+    // Destructor calls release()
+  }
+  VERIFY( test_mr.number_of_active_allocations() == 0 );
+}
+
+void
+test04()
+{
+  __gnu_test::memory_resource test_mr;
+  std::pmr::synchronized_pool_resource r({256, 256}, &test_mr);
+  // Check alignment
+  void* p1 = r.allocate(2, 64);
+  VERIFY( (std::uintptr_t)p1 % 64 == 0 );
+  void* p2 = r.allocate(2, 128);
+  VERIFY( (std::uintptr_t)p2 % 128 == 0 );
+  void* p3 = r.allocate(2, 256);
+  VERIFY( (std::uintptr_t)p3 % 256 == 0 );
+  const std::size_t largest_pool = r.options().largest_required_pool_block;
+  void* p4 = r.allocate(2 * largest_pool, 1024);
+  VERIFY( (std::uintptr_t)p4 % 1024 == 0 );
+  r.deallocate(p1, 2, 64);
+  r.deallocate(p2, 2, 128);
+  r.deallocate(p3, 2, 256);
+  r.deallocate(p4, 2 * largest_pool, 1024);
+}
+
+int
+main()
+{
+  test01();
+  test02();
+  test03();
+  test04();
+}
diff --git a/libstdc++-v3/testsuite/20_util/synchronized_pool_resource/cons.cc b/libstdc++-v3/testsuite/20_util/synchronized_pool_resource/cons.cc
new file mode 100644
index 00000000000..f02636ec1be
--- /dev/null
+++ b/libstdc++-v3/testsuite/20_util/synchronized_pool_resource/cons.cc
@@ -0,0 +1,83 @@
+// Copyright (C) 2018 Free Software Foundation, Inc.
+//
+// This file is part of the GNU ISO C++ Library.  This library is free
+// software; you can redistribute it and/or modify it under the
+// terms of the GNU General Public License as published by the
+// Free Software Foundation; either version 3, or (at your option)
+// any later version.
+
+// This library is distributed in the hope that it will be useful,
+// but WITHOUT ANY WARRANTY; without even the implied warranty of
+// MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+// GNU General Public License for more details.
+
+// You should have received a copy of the GNU General Public License along
+// with this library; see the file COPYING3.  If not see
+// <http://www.gnu.org/licenses/>.
+
+// { dg-do run }
+// { dg-options "-std=gnu++17 -pthread" }
+// { dg-require-effective-target c++17 }
+// { dg-require-effective-target pthread }
+// { dg-require-gthreads "" }
+
+#include <memory_resource>
+#include <testsuite_hooks.h>
+#include <testsuite_allocator.h>
+
+void
+test01()
+{
+  __gnu_test::memory_resource test_mr1, test_mr2;
+  __gnu_test::default_resource_mgr mgr(&test_mr1);
+
+  const std::pmr::pool_options opts{1, 2};
+  using std::pmr::synchronized_pool_resource;
+
+  synchronized_pool_resource p1 = {opts, &test_mr2};
+  VERIFY( p1.upstream_resource() == &test_mr2 );
+  synchronized_pool_resource p2;
+  VERIFY( p2.upstream_resource() == std::pmr::get_default_resource() );
+  synchronized_pool_resource p3{&test_mr2};
+  VERIFY( p3.upstream_resource() == &test_mr2 );
+  synchronized_pool_resource p4{opts};
+  VERIFY( p4.upstream_resource() == std::pmr::get_default_resource() );
+
+  static_assert(!std::is_copy_constructible_v<synchronized_pool_resource>);
+  static_assert(!std::is_copy_assignable_v<synchronized_pool_resource>);
+  static_assert(std::is_destructible_v<synchronized_pool_resource>);
+}
+
+void
+test02()
+{
+  __gnu_test::memory_resource test_mr1, test_mr2;
+  __gnu_test::default_resource_mgr mgr(&test_mr1);
+
+  const std::pmr::pool_options opts{1, 2};
+
+  struct derived : std::pmr::synchronized_pool_resource
+  {
+    using synchronized_pool_resource::synchronized_pool_resource;
+  };
+
+  derived p1 = {opts, &test_mr2};
+  VERIFY( p1.upstream_resource() == &test_mr2 );
+  derived p2;
+  VERIFY( p2.upstream_resource() == std::pmr::get_default_resource() );
+  derived p3{&test_mr2};
+  VERIFY( p3.upstream_resource() == &test_mr2 );
+  derived p4{opts};
+  VERIFY( p4.upstream_resource() == std::pmr::get_default_resource() );
+
+  static_assert(!std::is_copy_constructible_v<derived>);
+  static_assert(!std::is_copy_assignable_v<derived>);
+  static_assert(std::is_destructible_v<derived>);
+}
+
+int
+main()
+{
+  test01();
+  test02();
+}
diff --git a/libstdc++-v3/testsuite/20_util/synchronized_pool_resource/is_equal.cc b/libstdc++-v3/testsuite/20_util/synchronized_pool_resource/is_equal.cc
new file mode 100644
index 00000000000..68529387e9a
--- /dev/null
+++ b/libstdc++-v3/testsuite/20_util/synchronized_pool_resource/is_equal.cc
@@ -0,0 +1,41 @@
+// Copyright (C) 2018 Free Software Foundation, Inc.
+//
+// This file is part of the GNU ISO C++ Library.  This library is free
+// software; you can redistribute it and/or modify it under the
+// terms of the GNU General Public License as published by the
+// Free Software Foundation; either version 3, or (at your option)
+// any later version.
+
+// This library is distributed in the hope that it will be useful,
+// but WITHOUT ANY WARRANTY; without even the implied warranty of
+// MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+// GNU General Public License for more details.
+
+// You should have received a copy of the GNU General Public License along
+// with this library; see the file COPYING3.  If not see
+// <http://www.gnu.org/licenses/>.
+
+// { dg-do run }
+// { dg-options "-std=gnu++17 -pthread" }
+// { dg-require-effective-target c++17 }
+// { dg-require-effective-target pthread }
+// { dg-require-gthreads "" }
+
+#include <memory_resource>
+#include <testsuite_hooks.h>
+
+void
+test01()
+{
+  std::pmr::synchronized_pool_resource r1;
+  VERIFY( r1 == r1 );
+  std::pmr::synchronized_pool_resource r2;
+  VERIFY( r1 != r2 );
+  VERIFY( r2 != r1 );
+}
+
+int
+main()
+{
+  test01();
+}
diff --git a/libstdc++-v3/testsuite/20_util/synchronized_pool_resource/multithreaded.cc b/libstdc++-v3/testsuite/20_util/synchronized_pool_resource/multithreaded.cc
new file mode 100644
index 00000000000..d01d65a3bc6
--- /dev/null
+++ b/libstdc++-v3/testsuite/20_util/synchronized_pool_resource/multithreaded.cc
@@ -0,0 +1,86 @@
+// Copyright (C) 2018 Free Software Foundation, Inc.
+//
+// This file is part of the GNU ISO C++ Library.  This library is free
+// software; you can redistribute it and/or modify it under the
+// terms of the GNU General Public License as published by the
+// Free Software Foundation; either version 3, or (at your option)
+// any later version.
+
+// This library is distributed in the hope that it will be useful,
+// but WITHOUT ANY WARRANTY; without even the implied warranty of
+// MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+// GNU General Public License for more details.
+
+// You should have received a copy of the GNU General Public License along
+// with this library; see the file COPYING3.  If not see
+// <http://www.gnu.org/licenses/>.
+
+// { dg-do run }
+// { dg-options "-std=gnu++17 -pthread" }
+// { dg-require-effective-target c++17 }
+// { dg-require-effective-target pthread }
+// { dg-require-gthreads "" }
+
+#include <memory_resource>
+#include <future>
+#include <testsuite_allocator.h>
+#include <testsuite_hooks.h>
+
+void
+test01()
+{
+  __gnu_test::memory_resource test_mr;
+  std::pmr::synchronized_pool_resource smr(&test_mr);
+
+  const std::size_t largest_pool = smr.options().largest_required_pool_block;
+
+  auto do_alloc = [&smr](void*& p, size_t n) {
+    // perform some other allocations and deallocations on the same thread:
+    void* p2 = smr.allocate(n);
+    smr.deallocate(p2, n);
+    p2 = smr.allocate(n);
+    p = smr.allocate(n);
+    smr.deallocate(p2, n);
+  };
+  auto do_dealloc = [&smr](void* p, size_t n) { smr.deallocate(p, n); };
+
+  void* p1;
+  void* p2;
+  void* p3;
+  auto f1 = std::async(std::launch::async, do_alloc, std::ref(p1), 8);
+  auto f2 = std::async(std::launch::async, do_alloc, std::ref(p2), 64);
+  auto f3 = std::async(std::launch::async, do_alloc, std::ref(p3),
+		       largest_pool* 2);
+
+  f1.get();
+  f2.get();
+  f3.get();
+  VERIFY( p1 != nullptr );
+  VERIFY( p2 != nullptr );
+  VERIFY( p3 != nullptr );
+  size_t nallocs = test_mr.number_of_active_allocations();
+  VERIFY( nallocs >= 4 );
+
+  // deallocate on different threads from allocation:
+  f1 = std::async(std::launch::async, do_dealloc, p1, 8);
+  f2 = std::async(std::launch::async, do_dealloc, p2, 64);
+  f1.get();
+  f2.get();
+  // No additional memory is allocated by deallocating on new threads:
+  VERIFY( test_mr.number_of_active_allocations() == nallocs );
+
+  // Deallocate large unpooled allocation:
+  f3 = std::async(std::launch::async, do_dealloc, p3, largest_pool * 2);
+  f3.get();
+  // The large allocation should have been returned upstream:
+  VERIFY( test_mr.number_of_active_allocations() == nallocs - 1 );
+
+  smr.release();
+  VERIFY( test_mr.number_of_active_allocations() == 0 );
+}
+
+int
+main()
+{
+  test01();
+}
diff --git a/libstdc++-v3/testsuite/20_util/synchronized_pool_resource/options.cc b/libstdc++-v3/testsuite/20_util/synchronized_pool_resource/options.cc
new file mode 100644
index 00000000000..a0b63773d0c
--- /dev/null
+++ b/libstdc++-v3/testsuite/20_util/synchronized_pool_resource/options.cc
@@ -0,0 +1,54 @@
+// Copyright (C) 2018 Free Software Foundation, Inc.
+//
+// This file is part of the GNU ISO C++ Library.  This library is free
+// software; you can redistribute it and/or modify it under the
+// terms of the GNU General Public License as published by the
+// Free Software Foundation; either version 3, or (at your option)
+// any later version.
+
+// This library is distributed in the hope that it will be useful,
+// but WITHOUT ANY WARRANTY; without even the implied warranty of
+// MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+// GNU General Public License for more details.
+
+// You should have received a copy of the GNU General Public License along
+// with this library; see the file COPYING3.  If not see
+// <http://www.gnu.org/licenses/>.
+
+// { dg-do run }
+// { dg-options "-std=gnu++17 -pthread" }
+// { dg-require-effective-target c++17 }
+// { dg-require-effective-target pthread }
+// { dg-require-gthreads "" }
+
+#include <memory_resource>
+#include <testsuite_hooks.h>
+
+bool eq(const std::pmr::pool_options& lhs, const std::pmr::pool_options& rhs)
+{
+  return lhs.max_blocks_per_chunk == rhs.max_blocks_per_chunk
+    && lhs.largest_required_pool_block == rhs.largest_required_pool_block;
+}
+
+void
+test01()
+{
+  std::pmr::synchronized_pool_resource r0;
+  const std::pmr::pool_options opts = r0.options();
+  VERIFY( opts.max_blocks_per_chunk != 0 );
+  VERIFY( opts.largest_required_pool_block != 0 );
+
+  std::pmr::synchronized_pool_resource r1(opts);
+  const auto opts1 = r1.options();
+  VERIFY( eq(opts, opts1) );
+
+  std::pmr::synchronized_pool_resource r2(std::pmr::pool_options{0, 0});
+  const auto opts2 = r2.options();
+  VERIFY( eq(opts, opts2) );
+}
+
+int
+main()
+{
+  test01();
+}
diff --git a/libstdc++-v3/testsuite/20_util/synchronized_pool_resource/release.cc b/libstdc++-v3/testsuite/20_util/synchronized_pool_resource/release.cc
new file mode 100644
index 00000000000..af3055df564
--- /dev/null
+++ b/libstdc++-v3/testsuite/20_util/synchronized_pool_resource/release.cc
@@ -0,0 +1,116 @@
+// Copyright (C) 2018 Free Software Foundation, Inc.
+//
+// This file is part of the GNU ISO C++ Library.  This library is free
+// software; you can redistribute it and/or modify it under the
+// terms of the GNU General Public License as published by the
+// Free Software Foundation; either version 3, or (at your option)
+// any later version.
+
+// This library is distributed in the hope that it will be useful,
+// but WITHOUT ANY WARRANTY; without even the implied warranty of
+// MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+// GNU General Public License for more details.
+
+// You should have received a copy of the GNU General Public License along
+// with this library; see the file COPYING3.  If not see
+// <http://www.gnu.org/licenses/>.
+
+// { dg-do run }
+// { dg-options "-std=gnu++17 -pthread" }
+// { dg-require-effective-target c++17 }
+// { dg-require-effective-target pthread }
+// { dg-require-gthreads "" }
+
+#include <memory_resource>
+#include <testsuite_allocator.h>
+#include <testsuite_hooks.h>
+
+void
+test01()
+{
+  __gnu_test::memory_resource test_mr;
+  std::pmr::synchronized_pool_resource r(&test_mr);
+  r.release();
+  VERIFY( test_mr.number_of_active_allocations() == 0 );
+  r.release();
+  VERIFY( test_mr.number_of_active_allocations() == 0 );
+  (void) r.allocate(1);
+  VERIFY( test_mr.number_of_active_allocations() != 0 );
+  r.release();
+  VERIFY( test_mr.number_of_active_allocations() == 0 );
+  r.release();
+  VERIFY( test_mr.number_of_active_allocations() == 0 );
+}
+
+void
+test02()
+{
+  struct nullable_memory_resource : public std::pmr::memory_resource
+  {
+    void*
+    do_allocate(std::size_t bytes, std::size_t alignment) override
+    { return upstream->allocate(bytes, alignment); }
+
+    void
+    do_deallocate(void* p, std::size_t bytes, std::size_t alignment) override
+    { upstream->deallocate(p, bytes, alignment); }
+
+    bool
+    do_is_equal(const memory_resource& r) const noexcept override
+    { return &r == this; }
+
+    std::pmr::memory_resource* upstream = std::pmr::get_default_resource();
+  };
+
+  nullable_memory_resource test_mr;
+  std::pmr::synchronized_pool_resource r(&test_mr);
+  r.release();
+  test_mr.upstream = nullptr;
+  r.release(); // should not need to call anything through upstream pointer
+}
+
+void
+test03()
+{
+  __gnu_test::memory_resource test_mr;
+  {
+    std::pmr::synchronized_pool_resource r(&test_mr);
+    // Destructor calls release()
+  }
+  VERIFY( test_mr.number_of_active_allocations() == 0 );
+  {
+    std::pmr::synchronized_pool_resource r(&test_mr);
+    (void) r.allocate(1);
+    VERIFY( test_mr.number_of_active_allocations() != 0 );
+    // Destructor calls release()
+  }
+  VERIFY( test_mr.number_of_active_allocations() == 0 );
+  {
+    std::pmr::synchronized_pool_resource r({10, 16}, &test_mr);
+    (void) r.allocate(2 * r.options().largest_required_pool_block);
+    VERIFY( test_mr.number_of_active_allocations() != 0 );
+    // Destructor calls release()
+  }
+  VERIFY( test_mr.number_of_active_allocations() == 0 );
+  {
+    std::pmr::synchronized_pool_resource r({16, 16}, &test_mr);
+    (void) r.allocate(2);
+    (void) r.allocate(8);
+    (void) r.allocate(16);
+    (void) r.allocate(2);
+    (void) r.allocate(8);
+    (void) r.allocate(16);
+    (void) r.allocate(2 * r.options().largest_required_pool_block);
+    VERIFY( test_mr.number_of_active_allocations() != 0 );
+    // Destructor calls release()
+  }
+  VERIFY( test_mr.number_of_active_allocations() == 0 );
+}
+
+int
+main()
+{
+  test01();
+  test02();
+  test03();
+}
diff --git a/libstdc++-v3/testsuite/performance/20_util/memory_resource/pools.cc b/libstdc++-v3/testsuite/performance/20_util/memory_resource/pools.cc
index 77f37fe63b2..8a2a5f1facf 100644
--- a/libstdc++-v3/testsuite/performance/20_util/memory_resource/pools.cc
+++ b/libstdc++-v3/testsuite/performance/20_util/memory_resource/pools.cc
@@ -15,28 +15,27 @@
 
 // Override the -std flag in the check_performance script: STD=gnu++17
 
+// Run the test as both single- and multi-threaded: TEST_B
+
 #include <memory_resource>
 #include <list>
 #include <string>
 #include <testsuite_performance.h>
 
-struct size16 { char c[16]; };
-struct size32 { char c[32]; };
-struct size64 { char c[64]; };
-struct size128 { char c[128]; };
+const int iterations = 100;
 
 // Insert and remove elements of various sizes in std::list containers.
-// If report=true the function will measure and report the total performance
-// including the time taken to destroy the lists and deallocate everything.
-// If dest=false the function will measure and report the performance of
-// insert/remove operations only, not the destruction of the lists.
+// If timers!=nullptr the function will pause the timer while the lists
+// are cleared and deallocated, so that only insertions/removals are timed.
+// Otherwise, the time taken to deallocate the lists is also counted.
 void
-populate_lists(std::pmr::memory_resource* r, std::string name, bool dest,
-	       int kmax = 100)
+populate_lists(std::pmr::memory_resource* r, __gnu_test::time_counter* timers,
+	       int kmax = iterations)
 {
-  name += " std::list push/pop";
-  if (dest)
-    name += "/destroy";
+  struct size16 { char c[16]; };
+  struct size32 { char c[32]; };
+  struct size64 { char c[64]; };
+  struct size128 { char c[128]; };
 
   std::pmr::list<int> l4(r);
   std::pmr::list<size16> l16(r);
@@ -44,11 +43,6 @@ populate_lists(std::pmr::memory_resource* r, std::string name, bool dest,
   std::pmr::list<size64> l64(r);
   std::pmr::list<size128> l128(r);
 
-  using namespace __gnu_test;
-  time_counter time;
-  resource_counter resource;
-  start_counters(time, resource);
-
   const int imax = 1000;
   const int jmax = 100;
   for (int k = 0; k < kmax; ++k)
@@ -70,8 +64,8 @@ populate_lists(std::pmr::memory_resource* r, std::string name, bool dest,
       l128.pop_front();
     }
 
-    if (!dest)
-      time.stop();
+    if (timers)
+      timers->stop();
 
     // Deallocate everything:
     l4.clear();
@@ -80,35 +74,338 @@ populate_lists(std::pmr::memory_resource* r, std::string name, bool dest,
     l64.clear();
     l128.clear();
 
-    if (!dest)
-      time.restart();
+    if (timers)
+      timers->restart();
   }
-
-  stop_counters(time, resource);
-
-  report_performance(__FILE__, name.c_str(), time, resource);
-  clear_counters(time, resource);
 }
 
-int main()
+// Test allocations and deallocations of node-based containers (std::list).
+// In this test pmr::unsynchronized_pool_resource should be faster than
+// pmr::new_delete_resource().
+void test_lists_single_thread()
 {
   std::pmr::memory_resource* newdel = std::pmr::new_delete_resource();
   std::pmr::unsynchronized_pool_resource pool;
+#ifndef NOTHREAD
+  std::pmr::synchronized_pool_resource syncpool;
+#endif
 
-  for (auto b : { false, true })
+  auto run_test = [](auto* memres, std::string name, bool time_dtors) {
+    name += " std::list push/pop";
+    if (time_dtors)
+      name += "/destroy";
+    __gnu_test::time_counter time;
+    __gnu_test::resource_counter resource;
+    start_counters(time, resource);
+    populate_lists(memres, time_dtors ? nullptr : &time);
+    stop_counters(time, resource);
+    report_performance(__FILE__, name, time, resource);
+  };
+
+  for (auto time_dtors : {false, true})
   {
+    run_test(newdel, "new-delete-1 ", time_dtors);
+    run_test(newdel, "new-delete-2 ", time_dtors);
+    run_test(newdel, "new-delete-3 ", time_dtors);
+
     // Start with an empty set of pools:
     pool.release();
-
-    populate_lists(newdel, "new_delete 1", b);
-    populate_lists(newdel, "new_delete 2", b);
-    populate_lists(newdel, "new_delete 3", b);
-
-    populate_lists(&pool, "unsync pool 1", b);
+    run_test(&pool, "unsync-pool-1", time_dtors);
     // Destroy pools and start fresh:
     pool.release();
-    populate_lists(&pool, "unsync pool 2", b);
+    run_test(&pool, "unsync-pool-2", time_dtors);
     // Do not destroy pools, reuse allocated memory:
-    populate_lists(&pool, "unsync pool 3", b);
+    run_test(&pool, "unsync-pool-3", time_dtors);
+
+#ifndef NOTHREAD
+    syncpool.release();
+    run_test(&syncpool, "sync-pool-1  ", time_dtors);
+    // Destroy pools and start fresh:
+    syncpool.release();
+    run_test(&syncpool, "sync-pool-2  ", time_dtors);
+    // Do not destroy pools, reuse allocated memory:
+    run_test(&syncpool, "sync-pool-3  ", time_dtors);
+#endif
   }
 }
+
+// TODO test non-pooled large allocations from (un)synchronized_pool_resource
+
+#ifndef NOTHREAD
+# include <thread>
+# include <mutex>
+# include <cassert>
+
+// Multithreaded std::list test with each thread having its own resource.
+// (pmr::new_delete vs pmr::unsynchronized_pool vs pmr::synchronized_pool)
+//
+// In this test both pmr::unsynchronized_pool_resource and
+// pmr::synchronized_pool_resource should be faster than
+// pmr::new_delete_resource().
+void test_lists_resource_per_thread()
+{
+  std::mutex mx;
+  std::unique_lock<std::mutex> gate(mx, std::defer_lock);
+
+  struct state
+  {
+    std::thread thread;
+
+    // Per-thread pool resources:
+    std::pmr::unsynchronized_pool_resource unsync;
+    std::pmr::synchronized_pool_resource sync;
+
+    std::pmr::memory_resource* memres[3] = {
+      std::pmr::new_delete_resource(), &unsync, &sync
+    };
+  };
+
+  state states[4];
+
+  const std::string resnames[] = {"new-delete ", "unsync-pool", "sync-pool  "};
+
+  auto run_test = [&mx] (std::pmr::memory_resource* memres,
+			 __gnu_test::time_counter* timers)
+  {
+    std::lock_guard<std::mutex>{mx};  // block until the mutex can be locked
+    populate_lists(memres, timers);
+  };
+
+  auto time_threads = [&] (std::string testname, bool time_dtors, int which) {
+    __gnu_test::time_counter time;
+    __gnu_test::resource_counter resource;
+    gate.lock();
+    auto* time_ptr = time_dtors ? nullptr : &time;
+    for (auto& s : states)
+      s.thread = std::thread{ run_test, s.memres[which], time_ptr };
+    start_counters(time, resource);
+    gate.unlock(); // let the threads run
+    for (auto& s : states)
+      s.thread.join();
+    stop_counters(time, resource);
+    report_performance(__FILE__, resnames[which] + testname, time, resource);
+  };
+
+  for (auto time_dtors : {false, true})
+  {
+    std::string testname = " resource-per-thread std::list push/pop";
+    if (time_dtors)
+      testname += "/destroy";
+    for (int which : {0, 1, 2})
+      time_threads(testname, time_dtors, which);
+  }
+}
+
+// A naive memory_resource that adds a mutex to unsynchronized_pool_resource
+struct locking_pool_resource : std::pmr::unsynchronized_pool_resource
+{
+  void* do_allocate(std::size_t b, std::size_t a) override
+  {
+    std::lock_guard<std::mutex> l(m);
+    return unsynchronized_pool_resource::do_allocate(b, a);
+  }
+
+  void do_deallocate(void* p, std::size_t b, std::size_t a) override
+  {
+    std::lock_guard<std::mutex> l(m);
+    return unsynchronized_pool_resource::do_deallocate(p, b, a);
+  }
+
+  std::mutex m;
+};
+
+// Multithreaded std::list test with all threads sharing the same resource.
+// (new_delete vs unsynchronized_pool+mutex vs synchronized_pool)
+//
+// pmr::synchronized_pool_resource is not expected to be anywhere near
+// as fast as pmr::new_delete_resource() here, but should perform much
+// better than the naive locking_pool_resource type.
+void test_lists_shared_resource()
+{
+  std::mutex mx;
+  std::unique_lock<std::mutex> gate(mx, std::defer_lock);
+
+  locking_pool_resource unsync;
+  std::pmr::synchronized_pool_resource sync;
+
+  std::pmr::memory_resource* memres[3] = {
+    std::pmr::new_delete_resource(), &unsync, &sync
+  };
+
+  std::thread threads[4];
+
+  const std::string resnames[3] = { "new-delete", "mutex-pool", "sync-pool " };
+
+  auto run_test = [&mx] (std::pmr::memory_resource* memres,
+			 __gnu_test::time_counter* timers)
+  {
+    std::lock_guard<std::mutex>{mx};  // block until the mutex can be locked
+    populate_lists(memres, timers);
+  };
+
+  auto time_threads = [&] (std::string testname, bool time_dtors, int which) {
+    __gnu_test::time_counter time;
+    __gnu_test::resource_counter resource;
+    gate.lock();
+    auto* time_ptr = time_dtors ? nullptr : &time;
+    for (auto& t : threads)
+      t = std::thread{ run_test, memres[which], time_ptr };
+    start_counters(time, resource);
+    gate.unlock(); // let the threads run
+    for (auto& t : threads)
+      t.join();
+    stop_counters(time, resource);
+    report_performance(__FILE__, resnames[which] + testname, time, resource);
+  };
+
+  for (auto time_dtors : {false, true})
+  {
+    std::string testname = " shared-resource std::list push/pop";
+    if (time_dtors)
+      testname += "/destroy";
+    for (int which : {0, 1, 2})
+      time_threads(testname, time_dtors, which);
+  }
+}
+
+// TODO threaded test just doing loads of allocations, no deallocs
+// both with per-thread resource (unsync vs sync vs newdel)
+// and shared resource (locked vs sync vs newdel)
+
+// TODO threaded test just doing loads of deallocations, no allocs
+// both with per-thread resource (unsync vs sync vs newdel)
+// and shared resource (locked vs sync vs newdel)
+
+// Multithreaded test where deallocations happen on different threads.
+// (new_delete vs unsynchronized_pool+mutex vs synchronized_pool)
+//
+// This hits the slow path for pmr::synchronized_pool_resource, where
+// an exclusive lock must be taken to access other threads' pools.
+// pmr::synchronized_pool_resource is not expected to be anywhere near
+// as fast as pmr::new_delete_resource() here, but should perform much
+// better than the naive locking_pool_resource type.
+void test_cross_thread_dealloc()
+{
+  const int num_threads = 4;
+
+  struct X {
+    void* ptr;
+    unsigned size;
+  };
+
+  // A buffer for each thread, and extra buffers for half of the threads:
+  std::vector<X> allocs[num_threads * 3 / 2];
+  for (auto& v : allocs)
+    v.resize(1000 * iterations);
+
+  // Use a few different pools
+  const std::size_t sizes[] = { 8, 16, 8, 16, 32, 64, 8, 16, 32, 64 };
+
+  std::mutex mx;
+
+  auto run_test =
+  [&, num_threads] (std::pmr::memory_resource* memres, int i, bool with_exit)
+  {
+    std::size_t counter = 0;
+    std::lock_guard<std::mutex>{mx};
+    // Fill this thread's buffer with allocations:
+    for (X& x : allocs[i])
+    {
+      x.size = sizes[counter++ % 10];
+      x.ptr = memres->allocate(x.size, 1);
+    }
+
+    if (with_exit && i == 0)
+    {
+      // One of the threads exits, so that its pools transfer to the
+      // non-thread-specific list of pools.
+      return;
+    }
+    else if (i < num_threads / 2)
+    {
+      // Other threads continue allocating, into the extra buffers:
+      for (X& x : allocs[num_threads + i])
+      {
+	x.size = sizes[counter++ % 10];
+	x.ptr = memres->allocate(x.size, 1);
+      }
+    }
+    else
+    {
+      // Half of the threads start deallocating their own memory and the
+      // memory belonging to another pool
+      const int other = i - num_threads / 2;
+      for (unsigned n = 0; n < allocs[i].size(); ++n)
+      {
+	// Deallocate memory allocated in this thread:
+	X& x1 = allocs[i][n];
+	memres->deallocate(x1.ptr, x1.size, 1);
+	x1 = {};
+	// Deallocate memory allocated in another thread:
+	X& x2 = allocs[other][n];
+	memres->deallocate(x2.ptr, x2.size, 1);
+	x2 = {};
+      }
+    }
+  };
+
+  std::thread threads[num_threads];
+
+  locking_pool_resource unsync;
+  std::pmr::synchronized_pool_resource sync;
+
+  std::pmr::memory_resource* memres[3] = {
+    std::pmr::new_delete_resource(), &unsync, &sync
+  };
+  const std::string resnames[3] = { "new-delete", "mutex-pool", "sync-pool " };
+
+  auto time_threads = [&] (std::string name, int which, bool with_exit)
+  {
+    __gnu_test::time_counter time;
+    __gnu_test::resource_counter resource;
+    std::unique_lock<std::mutex> gate(mx);
+    for (auto& t : threads)
+      t = std::thread{ run_test, memres[which], &t - threads, with_exit };
+    start_counters(time, resource);
+    gate.unlock();
+    for (auto& t : threads)
+      t.join();
+    stop_counters(time, resource);
+    report_performance(__FILE__, resnames[which] + name, time, resource);
+
+    // Clean up:
+    for (auto& a : allocs)
+    {
+      const int i = (&a - allocs);
+      if (i < num_threads) // These allocations were freed
+	for (auto& x : a)
+	{
+	  assert(x.ptr == nullptr);
+	}
+      else if (with_exit && i == num_threads)
+	;
+      else
+	for (auto& x : a)
+	{
+	  memres[which]->deallocate(x.ptr, x.size, 1);
+	  x = {};
+	}
+    }
+  };
+
+  for (int which : {0, 1, 2})
+    time_threads(" cross-thread dealloc", which, false);
+  for (int which : {0, 1, 2})
+    time_threads(" cross-thread dealloc w/exit", which, true);
+}
+#endif
+
+int main()
+{
+  test_lists_single_thread();
+#ifndef NOTHREAD
+  test_lists_resource_per_thread();
+  test_lists_shared_resource();
+  test_cross_thread_dealloc();
+#endif
+}

Patch

diff --git a/libstdc++-v3/config/abi/pre/gnu.ver b/libstdc++-v3/config/abi/pre/gnu.ver
index 9d66f908e1a..c301fc31afd 100644
--- a/libstdc++-v3/config/abi/pre/gnu.ver
+++ b/libstdc++-v3/config/abi/pre/gnu.ver
@@ -2039,13 +2039,6 @@  GLIBCXX_3.4.26 {
     _ZNSt7__cxx1118basic_stringstreamI[cw]St11char_traitsI[cw]ESaI[cw]EEC[12]Ev;
     _ZNSt7__cxx1119basic_[io]stringstreamI[cw]St11char_traitsI[cw]ESaI[cw]EEC[12]Ev;
 
-    _ZNSt3pmr19new_delete_resourceEv;
-    _ZNSt3pmr20null_memory_resourceEv;
-    _ZNSt3pmr20get_default_resourceEv;
-    _ZNSt3pmr20set_default_resourceEPNS_15memory_resourceE;
-    _ZNSt3pmr25monotonic_buffer_resource13_M_new_bufferE[jmy][jmy];
-    _ZNSt3pmr25monotonic_buffer_resource18_M_release_buffersEv;
-
     # std::__throw_ios_failure(const char*, int);
     _ZSt19__throw_ios_failurePKci;
 
@@ -2057,6 +2050,18 @@  GLIBCXX_3.4.26 {
     _ZN11__gnu_debug25_Safe_local_iterator_base16_M_attach_singleEPNS_19_Safe_sequence_baseEb;
 
     # <memory_resource> members
+    _ZNSt3pmr19new_delete_resourceEv;
+    _ZNSt3pmr20null_memory_resourceEv;
+    _ZNSt3pmr20get_default_resourceEv;
+    _ZNSt3pmr20set_default_resourceEPNS_15memory_resourceE;
+    _ZNSt3pmr25monotonic_buffer_resource13_M_new_bufferE[jmy][jmy];
+    _ZNSt3pmr25monotonic_buffer_resource18_M_release_buffersEv;
+    _ZTINSt3pmr26synchronized_pool_resourceE;
+    _ZNSt3pmr26synchronized_pool_resourceC1ERKNS_12pool_optionsEPNS_15memory_resourceE;
+    _ZNSt3pmr26synchronized_pool_resourceD[12]Ev;
+    _ZNSt3pmr26synchronized_pool_resource7releaseEv;
+    _ZNSt3pmr26synchronized_pool_resource11do_allocateE[jmy][jmy];
+    _ZNSt3pmr26synchronized_pool_resource13do_deallocateEPv[jmy][jmy];
     _ZTINSt3pmr28unsynchronized_pool_resourceE;
     _ZNSt3pmr28unsynchronized_pool_resourceC[12]ERKNS_12pool_optionsEPNS_15memory_resourceE;
     _ZNSt3pmr28unsynchronized_pool_resourceD[12]Ev;
diff --git a/libstdc++-v3/include/std/memory_resource b/libstdc++-v3/include/std/memory_resource
index 40486af82fe..86866f6c4d5 100644
--- a/libstdc++-v3/include/std/memory_resource
+++ b/libstdc++-v3/include/std/memory_resource
@@ -37,6 +37,7 @@ 
 #include <utility>			// pair, index_sequence
 #include <vector>			// vector
 #include <cstddef>			// size_t, max_align_t
+#include <shared_mutex>			// shared_mutex
 #include <debug/assertions.h>
 
 namespace std _GLIBCXX_VISIBILITY(default)
@@ -338,7 +339,72 @@  namespace pmr
     const int _M_npools;
   };
 
-  // TODO class synchronized_pool_resource
+#ifdef _GLIBCXX_HAS_GTHREADS
+  /// A thread-safe memory resource that manages pools of fixed-size blocks.
+  class synchronized_pool_resource : public memory_resource
+  {
+  public:
+    synchronized_pool_resource(const pool_options& __opts,
+				 memory_resource* __upstream)
+    __attribute__((__nonnull__));
+
+    synchronized_pool_resource()
+    : synchronized_pool_resource(pool_options(), get_default_resource())
+    { }
+
+    explicit
+    synchronized_pool_resource(memory_resource* __upstream)
+    __attribute__((__nonnull__))
+    : synchronized_pool_resource(pool_options(), __upstream)
+    { }
+
+    explicit
+    synchronized_pool_resource(const pool_options& __opts)
+    : synchronized_pool_resource(__opts, get_default_resource()) { }
+
+    synchronized_pool_resource(const synchronized_pool_resource&) = delete;
+
+    virtual ~synchronized_pool_resource();
+
+    synchronized_pool_resource&
+    operator=(const synchronized_pool_resource&) = delete;
+
+    void release();
+
+    memory_resource*
+    upstream_resource() const noexcept
+    __attribute__((__returns_nonnull__))
+    { return _M_impl.resource(); }
+
+    pool_options options() const noexcept { return _M_impl._M_opts; }
+
+  protected:
+    void*
+    do_allocate(size_t __bytes, size_t __alignment) override;
+
+    void
+    do_deallocate(void* __p, size_t __bytes, size_t __alignment) override;
+
+    bool
+    do_is_equal(const memory_resource& __other) const noexcept override
+    { return this == &__other; }
+
+  public:
+    // Thread-specific pools (only public for access by implementation details)
+    struct _TPools;
+
+  private:
+    _TPools* _M_alloc_tpools(lock_guard<shared_mutex>&);
+    _TPools* _M_alloc_shared_tpools(lock_guard<shared_mutex>&);
+    auto _M_thread_specific_pools() noexcept;
+
+    __pool_resource _M_impl;
+    __gthread_key_t _M_key;
+    // Linked list of thread-specific pools. All threads share _M_tpools[0].
+    _TPools* _M_tpools = nullptr;
+    mutable shared_mutex _M_mx;
+  };
+#endif
 
   /// A non-thread-safe memory resource that manages pools of fixed-size blocks.
   class unsynchronized_pool_resource : public memory_resource
diff --git a/libstdc++-v3/include/std/shared_mutex b/libstdc++-v3/include/std/shared_mutex
index dce97f48a3f..8aa6e9f0d4f 100644
--- a/libstdc++-v3/include/std/shared_mutex
+++ b/libstdc++-v3/include/std/shared_mutex
@@ -57,6 +57,90 @@  _GLIBCXX_BEGIN_NAMESPACE_VERSION
   class shared_timed_mutex;
 
 #if _GLIBCXX_USE_PTHREAD_RWLOCK_T
+#ifdef __gthrw
+#define _GLIBCXX_GTHRW(name) \
+  __gthrw(pthread_ ## name); \
+  static inline int \
+  __glibcxx_ ## name (pthread_rwlock_t *__rwlock) \
+  { \
+    if (__gthread_active_p ()) \
+      return __gthrw_(pthread_ ## name) (__rwlock); \
+    else \
+      return 0; \
+  }
+  _GLIBCXX_GTHRW(rwlock_rdlock)
+  _GLIBCXX_GTHRW(rwlock_tryrdlock)
+  _GLIBCXX_GTHRW(rwlock_wrlock)
+  _GLIBCXX_GTHRW(rwlock_trywrlock)
+  _GLIBCXX_GTHRW(rwlock_unlock)
+# ifndef PTHREAD_RWLOCK_INITIALIZER
+  _GLIBCXX_GTHRW(rwlock_destroy)
+  __gthrw(pthread_rwlock_init);
+  static inline int
+  __glibcxx_rwlock_init (pthread_rwlock_t *__rwlock)
+  {
+    if (__gthread_active_p ())
+      return __gthrw_(pthread_rwlock_init) (__rwlock, NULL);
+    else
+      return 0;
+  }
+# endif
+# if _GTHREAD_USE_MUTEX_TIMEDLOCK
+   __gthrw(pthread_rwlock_timedrdlock);
+  static inline int
+  __glibcxx_rwlock_timedrdlock (pthread_rwlock_t *__rwlock,
+				const timespec *__ts)
+  {
+    if (__gthread_active_p ())
+      return __gthrw_(pthread_rwlock_timedrdlock) (__rwlock, __ts);
+    else
+      return 0;
+  }
+   __gthrw(pthread_rwlock_timedwrlock);
+  static inline int
+  __glibcxx_rwlock_timedwrlock (pthread_rwlock_t *__rwlock,
+				const timespec *__ts)
+  {
+    if (__gthread_active_p ())
+      return __gthrw_(pthread_rwlock_timedwrlock) (__rwlock, __ts);
+    else
+      return 0;
+  }
+# endif
+#else
+  static inline int
+  __glibcxx_rwlock_rdlock (pthread_rwlock_t *__rwlock)
+  { return pthread_rwlock_rdlock (__rwlock); }
+  static inline int
+  __glibcxx_rwlock_tryrdlock (pthread_rwlock_t *__rwlock)
+  { return pthread_rwlock_tryrdlock (__rwlock); }
+  static inline int
+  __glibcxx_rwlock_wrlock (pthread_rwlock_t *__rwlock)
+  { return pthread_rwlock_wrlock (__rwlock); }
+  static inline int
+  __glibcxx_rwlock_trywrlock (pthread_rwlock_t *__rwlock)
+  { return pthread_rwlock_trywrlock (__rwlock); }
+  static inline int
+  __glibcxx_rwlock_unlock (pthread_rwlock_t *__rwlock)
+  { return pthread_rwlock_unlock (__rwlock); }
+  static inline int
+  __glibcxx_rwlock_destroy(pthread_rwlock_t *__rwlock)
+  { return pthread_rwlock_destroy (__rwlock); }
+  static inline int
+  __glibcxx_rwlock_init(pthread_rwlock_t *__rwlock)
+  { return pthread_rwlock_init (__rwlock, NULL); }
+# if _GTHREAD_USE_MUTEX_TIMEDLOCK
+  static inline int
+  __glibcxx_rwlock_timedrdlock (pthread_rwlock_t *__rwlock,
+				const timespec *__ts)
+  { return pthread_rwlock_timedrdlock (__rwlock, __ts); }
+  static inline int
+  __glibcxx_rwlock_timedwrlock (pthread_rwlock_t *__rwlock,
+				const timespec *__ts)
+  { return pthread_rwlock_timedwrlock (__rwlock, __ts); }
+# endif
+#endif
+
   /// A shared mutex type implemented using pthread_rwlock_t.
   class __shared_mutex_pthread
   {
@@ -74,7 +158,7 @@  _GLIBCXX_BEGIN_NAMESPACE_VERSION
   public:
     __shared_mutex_pthread()
     {
-      int __ret = pthread_rwlock_init(&_M_rwlock, NULL);
+      int __ret = __glibcxx_rwlock_init(&_M_rwlock, NULL);
       if (__ret == ENOMEM)
 	__throw_bad_alloc();
       else if (__ret == EAGAIN)
@@ -87,7 +171,7 @@  _GLIBCXX_BEGIN_NAMESPACE_VERSION
 
     ~__shared_mutex_pthread()
     {
-      int __ret __attribute((__unused__)) = pthread_rwlock_destroy(&_M_rwlock);
+      int __ret __attribute((__unused__)) = __glibcxx_rwlock_destroy(&_M_rwlock);
       // Errors not handled: EBUSY, EINVAL
       __glibcxx_assert(__ret == 0);
     }
@@ -99,7 +183,7 @@  _GLIBCXX_BEGIN_NAMESPACE_VERSION
     void
     lock()
     {
-      int __ret = pthread_rwlock_wrlock(&_M_rwlock);
+      int __ret = __glibcxx_rwlock_wrlock(&_M_rwlock);
       if (__ret == EDEADLK)
 	__throw_system_error(int(errc::resource_deadlock_would_occur));
       // Errors not handled: EINVAL
@@ -109,7 +193,7 @@  _GLIBCXX_BEGIN_NAMESPACE_VERSION
     bool
     try_lock()
     {
-      int __ret = pthread_rwlock_trywrlock(&_M_rwlock);
+      int __ret = __glibcxx_rwlock_trywrlock(&_M_rwlock);
       if (__ret == EBUSY) return false;
       // Errors not handled: EINVAL
       __glibcxx_assert(__ret == 0);
@@ -119,7 +203,7 @@  _GLIBCXX_BEGIN_NAMESPACE_VERSION
     void
     unlock()
     {
-      int __ret __attribute((__unused__)) = pthread_rwlock_unlock(&_M_rwlock);
+      int __ret __attribute((__unused__)) = __glibcxx_rwlock_unlock(&_M_rwlock);
       // Errors not handled: EPERM, EBUSY, EINVAL
       __glibcxx_assert(__ret == 0);
     }
@@ -135,7 +219,7 @@  _GLIBCXX_BEGIN_NAMESPACE_VERSION
       // is okay based on the current specification of forward progress
       // guarantees by the standard.
       do
-	__ret = pthread_rwlock_rdlock(&_M_rwlock);
+	__ret = __glibcxx_rwlock_rdlock(&_M_rwlock);
       while (__ret == EAGAIN);
       if (__ret == EDEADLK)
 	__throw_system_error(int(errc::resource_deadlock_would_occur));
@@ -146,7 +230,7 @@  _GLIBCXX_BEGIN_NAMESPACE_VERSION
     bool
     try_lock_shared()
     {
-      int __ret = pthread_rwlock_tryrdlock(&_M_rwlock);
+      int __ret = __glibcxx_rwlock_tryrdlock(&_M_rwlock);
       // If the maximum number of read locks has been exceeded, we just fail
       // to acquire the lock.  Unlike for lock(), we are not allowed to throw
       // an exception.
@@ -413,7 +497,7 @@  _GLIBCXX_BEGIN_NAMESPACE_VERSION
 	    static_cast<long>(__ns.count())
 	  };
 
-	int __ret = pthread_rwlock_timedwrlock(&_M_rwlock, &__ts);
+	int __ret = __glibcxx_rwlock_timedwrlock(&_M_rwlock, &__ts);
 	// On self-deadlock, we just fail to acquire the lock.  Technically,
 	// the program violated the precondition.
 	if (__ret == ETIMEDOUT || __ret == EDEADLK)
@@ -466,7 +550,7 @@  _GLIBCXX_BEGIN_NAMESPACE_VERSION
 	// mistaken for a spurious failure, which might help users realise
 	// there is a deadlock.
 	do
-	  __ret = pthread_rwlock_timedrdlock(&_M_rwlock, &__ts);
+	  __ret = __glibcxx_rwlock_timedrdlock(&_M_rwlock, &__ts);
 	while (__ret == EAGAIN || __ret == EDEADLK);
 	if (__ret == ETIMEDOUT)
 	  return false;
diff --git a/libstdc++-v3/src/c++17/memory_resource.cc b/libstdc++-v3/src/c++17/memory_resource.cc
index 781bdada381..c34b817abe0 100644
--- a/libstdc++-v3/src/c++17/memory_resource.cc
+++ b/libstdc++-v3/src/c++17/memory_resource.cc
@@ -863,6 +863,11 @@  namespace pmr
     return 1;
   }
 
+#ifdef _GLIBCXX_HAS_GTHREADS
+  using shared_lock = std::shared_lock<shared_mutex>;
+  using exclusive_lock = lock_guard<shared_mutex>;
+#endif
+
   } // namespace
 
   __pool_resource::
@@ -948,6 +953,292 @@  namespace pmr
     return p;
   }
 
+#ifdef _GLIBCXX_HAS_GTHREADS
+  // synchronized_pool_resource members.
+
+  /* Notes on implementation and thread safety:
+   *
+   * Each synchronized_pool_resource manages an linked list of N+1 _TPools
+   * objects, where N is the number of threads using the pool resource.
+   * Each _TPools object has its own set of pools, with their own chunks.
+   * The first element of the list, _M_tpools[0], can be used by any thread.
+   * The rest of the list contains a _TPools object for each thread,
+   * accessed via the thread-specific key _M_key (and referred to for
+   * exposition as _M_tpools[_M_key]).
+   * The first element, _M_tpools[0], contains "orphaned chunks" which were
+   * allocated by a thread which has since exited, and so there is no
+   * _M_tpools[_M_key] for that thread.
+   * A thread can access its own thread-specific set of pools via _M_key
+   * while holding a shared lock on _M_mx. Accessing _M_impl._M_unpooled
+   * or _M_tpools[0] or any other thread's _M_tpools[_M_key] requires an
+   * exclusive lock.
+   * The upstream_resource() pointer can be obtained without a lock, but
+   * any dereference of that pointer requires an exclusive lock.
+   * The _M_impl._M_opts and _M_impl._M_npools members are immutable,
+   * and can safely be accessed concurrently.
+   */
+
+  extern "C" {
+    static void destroy_TPools(void*);
+  }
+
+  struct synchronized_pool_resource::_TPools
+  {
+    // Exclusive lock must be held in the thread where this constructor runs.
+    explicit
+    _TPools(synchronized_pool_resource& owner, exclusive_lock&)
+    : owner(owner), pools(owner._M_impl._M_alloc_pools())
+    {
+      // __builtin_printf("%p constructing\n", this);
+      __glibcxx_assert(pools);
+    }
+
+    // Exclusive lock must be held in the thread where this destructor runs.
+    ~_TPools()
+    {
+      __glibcxx_assert(pools);
+      if (pools)
+	{
+	  memory_resource* r = owner.upstream_resource();
+	  for (int i = 0; i < owner._M_impl._M_npools; ++i)
+	    pools[i].release(r);
+	  std::destroy_n(pools, owner._M_impl._M_npools);
+	  polymorphic_allocator<__pool_resource::_Pool> a(r);
+	  a.deallocate(pools, owner._M_impl._M_npools);
+	}
+      if (prev)
+	prev->next = next;
+      if (next)
+	next->prev = prev;
+    }
+
+    // Exclusive lock must be held in the thread where this function runs.
+    void move_nonempty_chunks()
+    {
+      __glibcxx_assert(pools);
+      if (!pools)
+	return;
+      memory_resource* r = owner.upstream_resource();
+      // move all non-empty chunks to the shared _TPools
+      for (int i = 0; i < owner._M_impl._M_npools; ++i)
+	for (auto& c : pools[i]._M_chunks)
+	  if (!c.empty())
+	    owner._M_tpools->pools[i]._M_chunks.insert(std::move(c), r);
+    }
+
+    synchronized_pool_resource& owner;
+    __pool_resource::_Pool* pools = nullptr;
+    _TPools* prev = nullptr;
+    _TPools* next = nullptr;
+
+    static void destroy(_TPools* p)
+    {
+      exclusive_lock l(p->owner._M_mx);
+      // __glibcxx_assert(p != p->owner._M_tpools);
+      p->move_nonempty_chunks();
+      polymorphic_allocator<_TPools> a(p->owner.upstream_resource());
+      p->~_TPools();
+      a.deallocate(p, 1);
+    }
+  };
+
+  // Called when a thread exits
+  extern "C" {
+    static void destroy_TPools(void* p)
+    {
+      using _TPools = synchronized_pool_resource::_TPools;
+      _TPools::destroy(static_cast<_TPools*>(p));
+    }
+  }
+
+  // Constructor
+  synchronized_pool_resource::
+  synchronized_pool_resource(const pool_options& opts,
+			     memory_resource* upstream)
+  : _M_impl(opts, upstream)
+  {
+    if (int err = __gthread_key_create(&_M_key, destroy_TPools))
+      __throw_system_error(err);
+    exclusive_lock l(_M_mx);
+    _M_tpools = _M_alloc_shared_tpools(l);
+  }
+
+  // Destructor
+  synchronized_pool_resource::~synchronized_pool_resource()
+  {
+    release();
+    __gthread_key_delete(_M_key); // does not run destroy_TPools
+  }
+
+  void
+  synchronized_pool_resource::release()
+  {
+    exclusive_lock l(_M_mx);
+    if (_M_tpools)
+      {
+	__gthread_key_delete(_M_key); // does not run destroy_TPools
+	__gthread_key_create(&_M_key, destroy_TPools);
+	polymorphic_allocator<_TPools> a(upstream_resource());
+	// destroy+deallocate each _TPools
+	do
+	  {
+	    _TPools* p = _M_tpools;
+	    _M_tpools = _M_tpools->next;
+	    p->~_TPools();
+	    a.deallocate(p, 1);
+	  }
+	while (_M_tpools);
+      }
+    // release unpooled memory
+    _M_impl.release();
+  }
+
+  // Caller must hold shared or exclusive lock to ensure the pointer
+  // isn't invalidated before it can be used.
+  auto
+  synchronized_pool_resource::_M_thread_specific_pools() noexcept
+  {
+    __pool_resource::_Pool* pools = nullptr;
+    if (auto tp = static_cast<_TPools*>(__gthread_getspecific(_M_key)))
+      {
+      pools = tp->pools;
+      __glibcxx_assert(tp->pools);
+      }
+    return pools;
+  }
+
+  // Override for memory_resource::do_allocate
+  void*
+  synchronized_pool_resource::
+  do_allocate(size_t bytes, size_t alignment)
+  {
+    const auto block_size = std::max(bytes, alignment);
+    if (block_size <= _M_impl._M_opts.largest_required_pool_block)
+      {
+	const ptrdiff_t index = pool_index(block_size, _M_impl._M_npools);
+	memory_resource* const r = upstream_resource();
+	const pool_options opts = _M_impl._M_opts;
+	{
+	  // Try to allocate from the thread-specific pool
+	  shared_lock l(_M_mx);
+	  if (auto pools = _M_thread_specific_pools()) // [[likely]]
+	    {
+	      // Need exclusive lock to replenish so use try_allocate:
+	      if (void* p = pools[index].try_allocate())
+		return p;
+	      // Need to take exclusive lock and replenish pool.
+	    }
+	  // Need to allocate or replenish thread-specific pools using
+	  // upstream resource, so need to hold exclusive lock.
+	}
+	// N.B. Another thread could call release() now lock is not held.
+	exclusive_lock excl(_M_mx);
+	if (!_M_tpools) // [[unlikely]]
+	  _M_tpools = _M_alloc_shared_tpools(excl);
+	auto pools = _M_thread_specific_pools();
+	if (!pools)
+	  pools = _M_alloc_tpools(excl)->pools;
+	return pools[index].allocate(r, opts);
+      }
+    exclusive_lock l(_M_mx);
+    return _M_impl.allocate(bytes, alignment); // unpooled allocation
+  }
+
+  // Override for memory_resource::do_deallocate
+  void
+  synchronized_pool_resource::
+  do_deallocate(void* p, size_t bytes, size_t alignment)
+  {
+    size_t block_size = std::max(bytes, alignment);
+    if (block_size <= _M_impl._M_opts.largest_required_pool_block)
+      {
+	const ptrdiff_t index = pool_index(block_size, _M_impl._M_npools);
+	__glibcxx_assert(index != -1);
+	{
+	  shared_lock l(_M_mx);
+	  auto pools = _M_thread_specific_pools();
+	  if (pools)
+	    {
+	      // No need to lock here, no other thread is accessing this pool.
+	      if (pools[index].deallocate(upstream_resource(), p))
+		return;
+	    }
+	  // Block might have come from a different thread's pool,
+	  // take exclusive lock and check every pool.
+	}
+	// TODO store {p, bytes, alignment} somewhere and defer returning
+	// the block to the correct thread-specific pool until we next
+	// take the exclusive lock.
+	exclusive_lock excl(_M_mx);
+	for (_TPools* t = _M_tpools; t != nullptr; t = t->next)
+	  {
+	    if (t->pools) // [[likely]]
+	      {
+		if (t->pools[index].deallocate(upstream_resource(), p))
+		  return;
+	      }
+	  }
+      }
+    exclusive_lock l(_M_mx);
+    _M_impl.deallocate(p, bytes, alignment);
+  }
+
+  // Allocate a thread-specific _TPools object and add it to the linked list.
+  auto
+  synchronized_pool_resource::_M_alloc_tpools(exclusive_lock& l)
+  -> _TPools*
+  {
+    __glibcxx_assert(_M_tpools != nullptr);
+    // dump_list(_M_tpools);
+    polymorphic_allocator<_TPools> a(upstream_resource());
+    _TPools* p = a.allocate(1);
+    bool constructed = false;
+    __try
+      {
+	a.construct(p, *this, l);
+	constructed = true;
+	// __glibcxx_assert(__gthread_getspecific(_M_key) == nullptr);
+	if (int err = __gthread_setspecific(_M_key, p))
+	  __throw_system_error(err);
+      }
+    __catch(...)
+      {
+	if (constructed)
+	  a.destroy(p);
+	a.deallocate(p, 1);
+	__throw_exception_again;
+      }
+    p->prev = _M_tpools;
+    p->next = _M_tpools->next;
+    _M_tpools->next = p;
+    if (p->next)
+      p->next->prev = p;
+    return p;
+  }
+
+  // Allocate the shared _TPools object, _M_tpools[0]
+  auto
+  synchronized_pool_resource::_M_alloc_shared_tpools(exclusive_lock& l)
+  -> _TPools*
+  {
+    __glibcxx_assert(_M_tpools == nullptr);
+    polymorphic_allocator<_TPools> a(upstream_resource());
+    _TPools* p = a.allocate(1);
+    __try
+      {
+	a.construct(p, *this, l);
+      }
+    __catch(...)
+      {
+	a.deallocate(p, 1);
+	__throw_exception_again;
+      }
+    // __glibcxx_assert(p->next == nullptr);
+    // __glibcxx_assert(p->prev == nullptr);
+    return p;
+  }
+#endif // _GLIBCXX_HAS_GTHREADS
+
   // unsynchronized_pool_resource member functions
 
   // Constructor