Delete INSTALL.md

Fixed compilation error on Windows while compiling with IntelLLVM compiler: "#include resolved using non-portable Microsoft search rules"

Signed-off-by: Ilya Isaev <ilya.isaev@intel.com>

.github/pull_request_template.md

@@ -0,0 +1,38 @@
+### Description 
+_Add comprehensive description of proposed changes_
+
+
+Fixes # - _issue number(s) if exists_
+
+### Type of change
+
+_Choose one or multiple, leave empty if none of the other choices apply_
+
+_Add respective label(s) to PR if you have permissions_
+
+- [ ] bug fix - _change which fixes an issue_
+- [ ] new feature - _change which adds functionality_
+- [ ] tests - _change in tests_
+- [ ] infrastructure - _change in infrastructure and CI_
+- [ ] documentation - _documentation update_
+
+### Tests
+
+- [ ] added - _required for new features and for some bug fixes_
+- [ ] not needed
+
+### Documentation
+
+- [ ] updated in # - _add PR number_
+- [ ] needs to be updated
+- [ ] not needed
+
+### Breaks backward compatibility
+- [ ] Yes
+- [ ] No
+- [ ] Unknown
+
+### Notify the following users
+_List users with `@` to send notifications_
+
+### Other information

CONTRIBUTING.md

@@ -38,7 +38,7 @@ You can use issues to report a problem, make a feature request, or add comments
 
 You can find all [open oneTBB pull requests](https://github.com/oneapi-src/oneTBB/pulls) on GitHub. 
 
-No anonymous contributions are accepted. The name in the commit message Signed-of-by line and your email must match the change authorship information.  Make sure your .gitconfig is set up correctly so you can use `git commit -s` for signing your patches: 
+No anonymous contributions are accepted. The name in the commit message Signed-off-by line and your email must match the change authorship information.  Make sure your .gitconfig is set up correctly so you can use `git commit -s` for signing your patches: 
 
 `git config --global user.name "Taylor Developer"`
 

INSTALL.md

@@ -1,34 +0,0 @@
-# Installation from Sources
-
-Required Software:
-* DPC++/C++ Compiler
-* Microsoft* Visual Studio* (Windows* OS only)
-* GNU Compilers (gcc) 4.8.5–10
-* GNU C Library (glibc) version 2.17–2.32
-* Clang* 4.2–11.0
-
-The project uses CMake build configuration. Make sure that you have CMake version 3.1 or higher. 
-
-Before installing the library, learn how to [configure, build, and test](cmake\README.md) the project. 
-
-NOTE: Be careful about installation. Avoid commands like `make install` unless you fully understand the consequences.
-
-The following installation components are supported:
-
-`runtime` - oneTBB runtime package (core shared libraries and .dll files on Windows).
-`devel` - oneTBB development package (header files, CMake integration files, library symbolic links, and .lib files on Windows).
-`tbb4py` - oneTBB Module for Python.
-
-To install the specific components after configuration and build, run:
-
-```bash
-cmake -DCOMPONENT=<component> [-DBUILD_TYPE=<build-type>] -P cmake_install.cmake
-```
-
-Simple packaging using CPack is supported.
-The following commands allow to create a simple portable package that includes header files, libraries, and integration files for CMake:
-
-```bash
-cmake <options> ..
-cpack
-```

README.md

@@ -5,8 +5,7 @@ oneAPI Threading Building Blocks (oneTBB) lets you easily write parallel C++ pro
 full advantage of multicore performance that are portable, composable, and have future-proof scalability.
 
 ## Installation
-You can install oneTBB from [GitHub*](https://github.com/oneapi-src/oneTBB/releases).
-See [Installation from Sources](INSTALL.md) for details.
+You can install oneTBB from GitHub*, see [Installation from Sources](INSTALL.md) for details.
 
 ## Documentation
 * [oneTBB Specification](https://spec.oneapi.com/versions/latest/elements/oneTBB/source/nested-index.html)
@@ -17,7 +16,7 @@ See [Installation from Sources](INSTALL.md) for details.
 
 ## Support
 Please report issues and suggestions via
-[GitHub issues](https://github.com/oneapi-src/oneTBB/issues).
+[GitHub issues](https://github.com/oneapi-src/oneTBB/issues). See our [documentation](./CONTRIBUTING.md##Issues) to learn how to work with issues. 
 
 ## How to Contribute
 We welcome community contributions, so check our [contributing guidelines](CONTRIBUTING.md)

cmake/compilers/MSVC.cmake

@@ -33,7 +33,7 @@ if (MSVC_VERSION LESS_EQUAL 1900)
     set(TBB_TEST_COMPILE_FLAGS ${TBB_TEST_COMPILE_FLAGS} /wd4503)
 endif()
 set(TBB_LIB_COMPILE_FLAGS -D_CRT_SECURE_NO_WARNINGS /GS)
-set(TBB_COMMON_COMPILE_FLAGS /volatile:iso /FS)
+set(TBB_COMMON_COMPILE_FLAGS /volatile:iso /FS /EHsc)
 
 # Ignore /WX set through add_compile_options() or added to CMAKE_CXX_FLAGS if TBB_STRICT is disabled.
 if (NOT TBB_STRICT AND COMMAND tbb_remove_compile_flag)

doc/GSG/index.rst

@@ -16,9 +16,18 @@ of multi-core processors. Use |short_name| to write scalable applications that:
 - Emphasize data parallel programming
 - Take advantage of concurrent collections and parallel algorithms
 
+System Requirements
+*******************
+
+Refer to the `oneTBB System Requirements <https://software.intel.com/content/www/us/en/develop/articles/intel-oneapi-threading-building-blocks-system-requirements.html>`_.
+
+
 Before You Begin
 ****************
 
+Download |short_name| as a `stand-alone product <https://software.intel.com/content/www/us/en/develop/articles/oneapi-standalone-components.html#onetbb>`_ 
+or as a part of the `Intel(R) oneAPI Base Toolkit <https://software.intel.com/content/www/us/en/develop/tools/oneapi/base-toolkit/download.html>`_.
+
 After installing |short_name|, you need to set the environment variables:
   
 #. Go to the oneTBB installation directory (``<install_dir>``). By default, ``<install_dir>`` is the following:
@@ -43,10 +52,25 @@ After installing |short_name|, you need to set the environment variables:
      ``vars.bat in <install_dir>/tbb/latest/env``
 
 
-System Requirements
-*******************
+Example
+*******
 
-Refer to the `oneTBB System Requirements <https://software.intel.com/content/www/us/en/develop/articles/intel-oneapi-threading-building-blocks-system-requirements.html>`_.
+Below you can find a typical example for a |short_name| algorithm. 
+The sample calculates a sum of all integer numbers from 1 to 100. 
+
+.. code:: cpp
+
+   int sum = oneapi::tbb::parallel_reduce(oneapi::tbb::blocked_range<int>(1,101), 0,
+      [](oneapi::tbb::blocked_range<int> const& r, int init) -> int {
+         for (int v = r.begin(); v != r.end(); v++  ) {
+            init += v;
+         }
+         return init;
+      },
+      [](int lhs, int rhs) -> int {
+         return lhs + rhs;
+      }
+   );
 
 Find more
 *********

doc/main/_static/custom.js

@@ -1,19 +1,3 @@
-//configure tms    
-var wapLocalCode = 'us-en'; 
-//dynamically set per localized site, see mapping table for values    
-var wapSection = "oneapi-docs";     
-//load tms    
-(function() {                 
-    var host = (window.document.location.protocol == 'http:') ? "http://www.intel.com" : "https://www.intel.com";
-    var url = host+"/content/dam/www/global/wap/tms-loader.js"; //wap file url
-    var po = document.createElement('script'); 
-    po.type = 'text/javascript'; 
-    po.async = true;  
-    po.src = url;                 
-    var s = document.getElementsByTagName('script')[0]; 
-    s.parentNode.insertBefore(po, s);     
-})(); 
-
 window.MathJax = {
     TeX: {
     Macros: {

doc/main/_templates/layout.html

@@ -0,0 +1,15 @@
+{% extends "!layout.html" %}
+{% block extrahead %}
+    <script defer type="text/javascript" src="https://www.intel.com/content/dam/www/global/wap/performance-config.js" ></script>
+    <script type="text/javascript">
+        // Configure TMS settings
+        var wapLocalCode = 'us-en'; // Dynamically set per localized site, see mapping table for values
+        var wapSection = "oneapi-tbb"; // WAP team will give you a unique section for your site
+        // Load TMS
+        (function () {
+          var url = 'https://www.intel.com/content/dam/www/global/wap/tms-loader.js'; // WAP file URL
+          var po = document.createElement('script'); po.type = 'text/javascript'; po.async = true; po.src = url;
+          var s = document.getElementsByTagName('script')[0]; s.parentNode.insertBefore(po, s);
+        })();
+      </script>
+{% endblock %}

doc/main/index.rst

@@ -1,6 +1,42 @@
 |full_name|
 ===========
 
+
+This document contains information about |full_name|. 
+|short_name| is a flexible performance library that can be found in the |base_tk|
+or as a stand-alone product. More information and specifications can be found on the
+|full_name| main page.
+
+Documentation for older versions of |short_name| is available for download only. 
+For a list of available documentation downloads by product version, see these pages:
+
+* `Download Documentation for Intel Parallel Studio XE <https://software.intel.com/content/www/us/en/develop/articles/download-documentation-intel-parallel-studio-xe-current-previous.html>`_
+* `Download Documentation for Intel System Studio <https://software.intel.com/content/www/us/en/develop/articles/download-documentation-intel-system-studio-current-previous.html>`_
+
+The following are some important topics for the ``novice user``:
+
+**Get Started with oneAPI Threading Building Blocks (oneTBB)**
+
+`Get Started with oneTBB <https://software.intel.com/content/www/us/en/develop/documentation/get-started-with-onetbb/top.html>`_ gives you a brief explanation of what oneTBB is. 
+
+**oneTBB Benefits**
+
+:ref:`Benefits` describes how |short_name| differs from typical threading packages.
+
+**Package Contents**
+
+|short_name| includes dynamic library files and header files for Windows*, Linux*, and macOS* 
+operating systems as described in :ref:`Package_Contents`.
+
+
+The following is important topic for the ``experienced user``:
+
+**Migrating from Threading Building Blocks (TBB)**
+
+Learn how to migrate from TBB to oneTBB with :ref:`Migration_Guide`.
+
+
+
 .. toctree::
    :maxdepth: 2
 
@@ -10,4 +46,4 @@
    intro/Benefits
    tbb_userguide/title
    reference/reference
-   intro/notices_and_disclaimers
+   intro/notices_and_disclaimers

doc/main/tbb_userguide/Migration_Guide/Task_API.rst

@@ -325,3 +325,131 @@ task:
         }
         tg.wait();
     }
+
+
+Scheduler Bypass
+----------------
+
+TBB ``task::execute()`` method can return a pointer to a task that can be executed next by the current thread.
+This might reduce scheduling overheads compared to direct ``spawn``. Similar to ``spawn``, the returned task 
+is not guaranteed to be executed next by the current thread.
+
+.. code:: cpp
+    
+    #include <tbb/task.h>
+    
+    // Assuming OtherTask is defined.
+    
+    struct Task : tbb::task {
+        task* execute(){
+            // some work to do ...
+            
+            auto* other_p = new(this->parent().allocate_child()) OtherTask{};
+            this->parent().add_ref_count();
+            
+            return other_p;
+        }
+    };
+    
+    int main(){
+        // Assuming RootTask is  defined.
+        RootTask& root = *new(tbb::task::allocate_root()) RootTask{};
+    
+        Task& child = *new(root.allocate_child()) Task{/*params*/};
+        
+        root.add_ref_count();
+        
+        tbb::task_spawn(child);
+        
+        root.wait_for_all();;
+    }
+
+In oneTBB this can be done using the preview feature of ``oneapi::tbb::task_group``. 
+
+.. code:: cpp
+   
+    #define TBB_PREVIEW_TASK_GROUP_EXTENSIONS 1
+    #include <oneapi/tbb/task_group.h>
+    
+    // Assuming OtherTask is defined.
+    
+    int main(){
+        oneapi::tbb::task_group tg;
+        
+        tg.run([&tg](){
+            //some work to do ...
+            
+            return tg.defer(OtherTask{});
+        });
+        
+        tg.wait();
+    }
+
+Here ``oneapi::tbb::task_group::defer`` adds a new task into the ``tg``. However, the task is not put into a 
+queue of tasks ready for execution via ``oneapi::tbb::task_group::run``, but bypassed to the executing thread directly 
+via function return value. 
+
+Deferred task creation
+----------------------
+The TBB low-level task API separates the task creation from the actual spawning. This separation allows to
+postpone the task spawning, while the parent task and final result production are blocked from premature leave. 
+For example, ``RootTask``, ``ChildTask``, and ``CallBackTask`` are the user-side functors that
+inherit ``tbb::task`` and implement its interface. Then, blocking the ``RootTask`` from leaving prematurely
+and waiting on it is implemented as follows: 
+
+.. code:: cpp
+
+    #include <tbb/task.h>
+
+    int main() {
+        // Assuming RootTask, ChildTask, and CallBackTask are defined.
+        RootTask& root = *new(tbb::task::allocate_root()) RootTask{};
+
+        ChildTask&    child    = *new(root.allocate_child()) ChildTask{/*params*/};
+        CallBackTask& cb_task  = *new(root.allocate_child()) CallBackTask{/*params*/};
+
+        root.set_ref_count(3);
+        
+        tbb::task::spawn(child);
+        
+        register_callback([cb_task&](){
+            tbb::task::enqueue(cb_task);
+        });
+
+        root.wait_for_all();
+        // Control flow will reach here only after both ChildTask and CallBackTask are executed,
+        // i.e. after the callback is called  
+    }
+
+In oneTBB this can be done using the preview feature of ``oneapi::tbb::task_group``.
+
+.. code:: cpp
+
+    #define TBB_PREVIEW_TASK_GROUP_EXTENSIONS 1
+    #include <oneapi/tbb/task_group.h>
+
+    int main(){
+        oneapi::tbb::task_group tg;
+        oneapi::tbb::task_arena arena;
+        // Assuming ChildTask and CallBackTask are defined.
+        
+        auto cb = tg.defer(CallBackTask{/*params*/});
+        
+        register_callback([&tg, c = std::move(cb), &arena]{
+            arena.enqueue(c);
+        });        
+
+        tg.run(ChildTask{/*params*/});
+
+ 
+        tg.wait();
+        // Control flow gets here once both ChildTask and CallBackTask are executed
+        // i.e. after the callback is called  
+    }
+
+Here ``oneapi::tbb::task_group::defer`` adds a new task into the ``tg``. However, the task is not spawned until 
+``oneapi::tbb::task_arena::enqueue`` is called. 
+
+.. note::
+   The call to ``oneapi::tbb::task_group::wait`` will not return control until both ``ChildTask`` and 
+   ``CallBackTask`` are executed.

include/oneapi/tbb/collaborative_call_once.h

@@ -19,9 +19,9 @@
 
 #if __TBB_PREVIEW_COLLABORATIVE_CALL_ONCE
 
-#include "tbb/task_arena.h"
-#include "tbb/task_group.h"
-#include "tbb/task.h"
+#include "task_arena.h"
+#include "task_group.h"
+#include "task.h"
 
 #include <atomic>
 

include/oneapi/tbb/task_group.h

@@ -524,6 +524,28 @@ protected:
         return cancellation_status ? canceled : complete;
     }
 
+#if __TBB_PREVIEW_TASK_GROUP_EXTENSIONS
+    task_group_status internal_run_and_wait(d2::task_handle&& h) {
+        if (h == nullptr) {
+            throw_exception(exception_id::bad_task_handle);
+        }
+
+        using acs = d2::task_handle_accessor;
+        if (&acs::ctx_of(h) != &context()) {
+            throw_exception(exception_id::bad_task_handle_wrong_task_group);
+        }
+
+        bool cancellation_status = false;
+        try_call([&] {
+            execute_and_wait(*acs::release(h), context(), m_wait_ctx, context());
+        }).on_completion([&] {
+            // TODO: the reset method is not thread-safe. Ensure the correct behavior.
+            cancellation_status = context().is_group_execution_cancelled();
+            context().reset();
+        });
+        return cancellation_status ? canceled : complete;
+    }
+#endif
     template<typename F>
     task* prepare_task(F&& f) {
         m_wait_ctx.reserve();
@@ -632,6 +654,12 @@ public:
     task_group_status run_and_wait(const F& f) {
         return internal_run_and_wait(f);
     }
+
+#if __TBB_PREVIEW_TASK_GROUP_EXTENSIONS
+    task_group_status run_and_wait(d2::task_handle&& h) {
+        return internal_run_and_wait(std::move(h));
+    }
+#endif
 }; // class task_group
 
 #if TBB_PREVIEW_ISOLATED_TASK_GROUP

include/oneapi/tbb/version.h

@@ -23,7 +23,7 @@
 // Product version
 #define TBB_VERSION_MAJOR 2021
 // Update version
-#define TBB_VERSION_MINOR 4
+#define TBB_VERSION_MINOR 5
 // "Patch" version for custom releases
 #define TBB_VERSION_PATCH 0
 // Suffix string
@@ -34,7 +34,7 @@
 // OneAPI oneTBB specification version
 #define ONETBB_SPEC_VERSION "1.0"
 // Full interface version
-#define TBB_INTERFACE_VERSION 12040
+#define TBB_INTERFACE_VERSION 12050
 // Major interface version
 #define TBB_INTERFACE_VERSION_MAJOR (TBB_INTERFACE_VERSION/1000)
 // Minor interface version

src/tbb/arena.h

@@ -217,9 +217,6 @@ struct arena_base : padded<intrusive_list_node> {
     task_stream<back_nonnull_accessor> my_critical_task_stream;
 #endif
 
-    //! The number of workers requested by the external thread owning the arena.
-    unsigned my_max_num_workers;
-
     //! The total number of workers that are requested from the resource manager.
     int my_total_num_workers_requested;
 
@@ -254,18 +251,9 @@ struct arena_base : padded<intrusive_list_node> {
     //! The market that owns this arena.
     market* my_market;
 
-    //! ABA prevention marker.
-    std::uintptr_t my_aba_epoch;
-
     //! Default task group context.
     d1::task_group_context* my_default_ctx;
 
-    //! The number of slots in the arena.
-    unsigned my_num_slots;
-
-    //! The number of reserved slots (can be occupied only by external threads).
-    unsigned my_num_reserved_slots;
-
 #if __TBB_ENQUEUE_ENFORCED_CONCURRENCY
     // arena needs an extra worker despite a global limit
     std::atomic<bool> my_global_concurrency_mode;
@@ -284,6 +272,21 @@ struct arena_base : padded<intrusive_list_node> {
     int my_local_concurrency_requests;
 #endif /* __TBB_ENQUEUE_ENFORCED_CONCURRENCY*/
 
+    //! ABA prevention marker.
+    std::uintptr_t my_aba_epoch;
+    //! The number of slots in the arena.
+    unsigned my_num_slots;
+    //! The number of reserved slots (can be occupied only by external threads).
+    unsigned my_num_reserved_slots;
+    //! The number of workers requested by the external thread owning the arena.
+    unsigned my_max_num_workers;
+
+    //! The target serialization epoch for callers of adjust_job_count_estimate
+    int my_adjust_demand_target_epoch;
+
+    //! The current serialization epoch for callers of adjust_job_count_estimate
+    d1::waitable_atomic<int> my_adjust_demand_current_epoch;
+
 #if TBB_USE_ASSERT
     //! Used to trap accesses to the object after its destruction.
     std::uintptr_t my_guard;

src/tbb/market.cpp

@@ -578,14 +578,14 @@ void market::adjust_demand ( arena& a, int delta, bool mandatory ) {
         my_num_workers_requested += delta;
         __TBB_ASSERT(my_num_workers_requested <= (int)effective_soft_limit, NULL);
 
-        target_epoch = my_adjust_demand_target_epoch++;
+        target_epoch = a.my_adjust_demand_target_epoch++;
     }
 
-    my_adjust_demand_current_epoch.wait_until(target_epoch, /* context = */ target_epoch, std::memory_order_acquire);
+    a.my_adjust_demand_current_epoch.wait_until(target_epoch, /* context = */ target_epoch, std::memory_order_relaxed);
     // Must be called outside of any locks
     my_server->adjust_job_count_estimate( delta );
-    my_adjust_demand_current_epoch.exchange(target_epoch + 1);
-    my_adjust_demand_current_epoch.notify_relaxed(target_epoch + 1);
+    a.my_adjust_demand_current_epoch.exchange(target_epoch + 1);
+    a.my_adjust_demand_current_epoch.notify_relaxed(target_epoch + 1);
 }
 
 void market::process( job& j ) {

src/tbb/market.h

@@ -105,12 +105,6 @@ private:
     //! Number of workers currently requested from RML
     int my_num_workers_requested;
 
-    //! The target serialization epoch for callers of adjust_job_count_estimate
-    int my_adjust_demand_target_epoch;
-
-    //! The current serialization epoch for callers of adjust_job_count_estimate
-    d1::waitable_atomic<int> my_adjust_demand_current_epoch;
-
     //! First unused index of worker
     /** Used to assign indices to the new workers coming from RML, and busy part
         of my_workers array. **/

src/tbb/queuing_rw_mutex.cpp

@@ -168,7 +168,11 @@ struct queuing_rw_mutex_impl {
         s.my_state.store(d1::queuing_rw_mutex::scoped_lock::state_t(write ? STATE_WRITER : STATE_READER), std::memory_order_relaxed);
         s.my_internal_lock.store(RELEASED, std::memory_order_relaxed);
 
-        queuing_rw_mutex::scoped_lock* predecessor = m.q_tail.exchange(&s, std::memory_order_release);
+
+        // The CAS must have release semantics, because we are
+        // "sending" the fields initialized above to other actors.
+        // We need acquire semantics, because we are acquiring the predecessor (or mutex if no predecessor)
+        queuing_rw_mutex::scoped_lock* predecessor = m.q_tail.exchange(&s, std::memory_order_acq_rel);
 
         if( write ) {       // Acquiring for write
 
@@ -177,7 +181,8 @@ struct queuing_rw_mutex_impl {
                 predecessor = tricky_pointer(predecessor) & ~FLAG;
                 __TBB_ASSERT( !predecessor->my_next, "the predecessor has another successor!");
                 tricky_pointer::store(predecessor->my_next, &s, std::memory_order_release);
-                spin_wait_until_eq(s.my_going, 1U);
+                // We are acquiring the mutex
+                spin_wait_until_eq(s.my_going, 1U, std::memory_order_acquire);
             }
 
         } else {            // Acquiring for read
@@ -194,10 +199,18 @@ struct queuing_rw_mutex_impl {
                 } else {
                     // Load predecessor->my_state now, because once predecessor->my_next becomes
                     // non-NULL, we must assume that *predecessor might be destroyed.
-                    pred_state = STATE_READER;
-                    predecessor->my_state.compare_exchange_strong(pred_state, STATE_READER_UNBLOCKNEXT, std::memory_order_acq_rel);
+                    pred_state = predecessor->my_state.load(std::memory_order_relaxed);
+                    if (pred_state == STATE_READER) {
+                        // Notify the previous reader to unblock us.
+                        predecessor->my_state.compare_exchange_strong(pred_state, STATE_READER_UNBLOCKNEXT, std::memory_order_relaxed);
+                    }
+                    if (pred_state == STATE_ACTIVEREADER)  { // either we initially read it or CAS failed
+                        // Active reader means that the predecessor already acquired the mutex and cannot notify us.
+                        // Therefore, we need to acquire the mutex ourselves by re-reading predecessor state.
+                        (void)predecessor->my_state.load(std::memory_order_acquire);
+                    }
                 }
-                tricky_pointer::store(s.my_prev, predecessor, std::memory_order_relaxed); // TODO: order?
+                tricky_pointer::store(s.my_prev, predecessor, std::memory_order_relaxed);
                 __TBB_ASSERT( !( tricky_pointer(predecessor) & FLAG ), "use of corrupted pointer!" );
                 __TBB_ASSERT( !predecessor->my_next.load(std::memory_order_relaxed), "the predecessor has another successor!");
                 tricky_pointer::store(predecessor->my_next, &s, std::memory_order_release);
@@ -206,13 +219,18 @@ struct queuing_rw_mutex_impl {
                     sync_prepare_done = true;
                     ITT_NOTIFY(sync_prepare, s.my_mutex);
     #endif
+                    // We are acquiring the mutex
                     spin_wait_until_eq(s.my_going, 1U, std::memory_order_acquire);
                 }
             }
 
             // The protected state must have been acquired here before it can be further released to any other reader(s):
             unsigned char old_state = STATE_READER;
-            s.my_state.compare_exchange_strong(old_state, STATE_ACTIVEREADER, std::memory_order_acq_rel);
+            // When this reader is signaled by previous actor it acquires the mutex.
+            // We need to build happens-before relation with all other coming readers that will read our ACTIVEREADER
+            // without blocking on my_going. Therefore, we need to publish ACTIVEREADER with release semantics.
+            // On fail it is relaxed, because we will build happens-before on my_going.
+            s.my_state.compare_exchange_strong(old_state, STATE_ACTIVEREADER, std::memory_order_release, std::memory_order_relaxed);
             if( old_state!=STATE_READER ) {
 #if __TBB_USE_ITT_NOTIFY
                 if( !sync_prepare_done )
@@ -220,20 +238,16 @@ struct queuing_rw_mutex_impl {
 #endif
                 // Failed to become active reader -> need to unblock the next waiting reader first
                 __TBB_ASSERT( s.my_state.load(std::memory_order_relaxed)==STATE_READER_UNBLOCKNEXT, "unexpected state" );
-                spin_wait_while_eq(s.my_next, 0U);
+                spin_wait_while_eq(s.my_next, 0U, std::memory_order_acquire);
                 /* my_state should be changed before unblocking the next otherwise it might finish
                    and another thread can get our old state and left blocked */
                 s.my_state.store(STATE_ACTIVEREADER, std::memory_order_relaxed);
                 tricky_pointer::load(s.my_next, std::memory_order_relaxed)->my_going.store(1U, std::memory_order_release);
             }
-            __TBB_ASSERT( s.my_state.load(std::memory_order_relaxed) ==STATE_ACTIVEREADER, "unlocked reader is active reader" );
+            __TBB_ASSERT(s.my_state.load(std::memory_order_relaxed) == STATE_ACTIVEREADER, "unlocked reader is active reader");
         }
 
         ITT_NOTIFY(sync_acquired, s.my_mutex);
-
-        // Force acquire so that user's critical section receives correct values
-        // from processor that was previously in the user's critical section.
-        atomic_fence(std::memory_order_acquire); // TODO: consider exchange on line 171
     }
 
     //! A method to acquire queuing_rw_mutex if it is free
@@ -253,13 +267,11 @@ struct queuing_rw_mutex_impl {
         s.my_internal_lock.store(RELEASED, std::memory_order_relaxed);
 
         // The CAS must have release semantics, because we are
-        // "sending" the fields initialized above to other processors.
+        // "sending" the fields initialized above to other actors.
+        // We need acquire semantics, because we are acquiring the mutex
         d1::queuing_rw_mutex::scoped_lock* expected = nullptr;
-        if (!m.q_tail.compare_exchange_strong(expected, &s, std::memory_order_release))
+        if (!m.q_tail.compare_exchange_strong(expected, &s, std::memory_order_acq_rel))
             return false; // Someone already took the lock
-        // Force acquire so that user's critical section receives correct values
-        // from processor that was previously in the user's critical section.
-        atomic_fence(std::memory_order_acquire);
         s.my_mutex = &m;
         ITT_NOTIFY(sync_acquired, s.my_mutex);
         return true;
@@ -280,32 +292,50 @@ struct queuing_rw_mutex_impl {
             d1::queuing_rw_mutex::scoped_lock* next = tricky_pointer::load(s.my_next, std::memory_order_acquire);
             if( !next ) {
                 d1::queuing_rw_mutex::scoped_lock* expected = &s;
-                if( s.my_mutex->q_tail.compare_exchange_strong(expected, nullptr, std::memory_order_release) ) {
+                // Release mutex on success otherwise wait for successor publication
+                if( s.my_mutex->q_tail.compare_exchange_strong(expected, nullptr,
+                    std::memory_order_release, std::memory_order_relaxed) )
+                {
                     // this was the only item in the queue, and the queue is now empty.
                     goto done;
                 }
-                spin_wait_while_eq(s.my_next, 0U);
+                spin_wait_while_eq(s.my_next, 0U, std::memory_order_relaxed);
                 next = tricky_pointer::load(s.my_next, std::memory_order_acquire);
             }
             next->my_going.store(2U, std::memory_order_relaxed); // protect next queue node from being destroyed too early
-            if( next->my_state.load()==STATE_UPGRADE_WAITING ) {
+            // If the next is STATE_UPGRADE_WAITING, it is expected to acquire all other released readers via release
+            // sequence in next->my_state. In that case, we need to preserve release sequence in next->my_state
+            // contributed by other reader. So, there are two approaches not to break the release sequence:
+            //   1. Use read-modify-write (exchange) operation to store with release the UPGRADE_LOSER state;
+            //   2. Acquire the release sequence and store the sequence and UPGRADE_LOSER state.
+            // The second approach seems better on x86 because it does not involve interlocked operations.
+            // Therefore, we read next->my_state with acquire while it is not required for else branch to get the 
+            // release sequence.
+            if( next->my_state.load(std::memory_order_acquire)==STATE_UPGRADE_WAITING ) {
                 // the next waiting for upgrade means this writer was upgraded before.
                 acquire_internal_lock(s);
                 // Responsibility transition, the one who reads uncorrupted my_prev will do release.
+                // Guarantee that above store of 2 into next->my_going happens-before resetting of next->my_prev
                 d1::queuing_rw_mutex::scoped_lock* tmp = tricky_pointer::exchange(next->my_prev, nullptr, std::memory_order_release);
-                next->my_state.store(STATE_UPGRADE_LOSER, std::memory_order_relaxed);
+                // Pass the release sequence that we acquired with the above load of next->my_state.
+                next->my_state.store(STATE_UPGRADE_LOSER, std::memory_order_release);
+                // We are releasing the mutex
                 next->my_going.store(1U, std::memory_order_release);
                 unblock_or_wait_on_internal_lock(s, get_flag(tmp));
             } else {
                 // next->state cannot be STATE_UPGRADE_REQUESTED
                 __TBB_ASSERT( next->my_state.load(std::memory_order_relaxed) & (STATE_COMBINED_WAITINGREADER | STATE_WRITER), "unexpected state" );
                 __TBB_ASSERT( !( next->my_prev.load(std::memory_order_relaxed) & FLAG ), "use of corrupted pointer!" );
-                tricky_pointer::store(next->my_prev, nullptr, std::memory_order_relaxed);
+                // Guarantee that above store of 2 into next->my_going happens-before resetting of next->my_prev
+                tricky_pointer::store(next->my_prev, nullptr, std::memory_order_release);
+                // We are releasing the mutex
                 next->my_going.store(1U, std::memory_order_release);
             }
 
         } else { // Acquired for read
-
+            // The basic idea it to build happens-before relation with left and right readers via prev and next. In addition,
+            // the first reader should acquire the left (prev) signal and propagate to right (next). To simplify, we always
+            // build happens-before relation between left and right (left is happened before right).
             queuing_rw_mutex::scoped_lock *tmp = nullptr;
     retry:
             // Addition to the original paper: Mark my_prev as in use
@@ -317,12 +347,9 @@ struct queuing_rw_mutex_impl {
                     // Failed to acquire the lock on predecessor. The predecessor either unlinks or upgrades.
                     // In the second case, it could or could not know my "in use" flag - need to check
                     // Responsibility transition, the one who reads uncorrupted my_prev will do release.
-                    tmp = tricky_pointer::compare_exchange_strong(s.my_prev, tricky_pointer(predecessor) | FLAG, predecessor, std::memory_order_release);
+                    tmp = tricky_pointer::compare_exchange_strong(s.my_prev, tricky_pointer(predecessor) | FLAG, predecessor, std::memory_order_acquire);
                     if( !(tricky_pointer(tmp) & FLAG) ) {
-                        // Wait for the predecessor to change my_prev (e.g. during unlink)
-                        // TODO: spin_wait condition seems never reachable
                         __TBB_ASSERT(tricky_pointer::load(s.my_prev, std::memory_order_relaxed) != (tricky_pointer(predecessor) | FLAG), nullptr);
-                        tricky_pointer::spin_wait_while_eq( s.my_prev, tricky_pointer(predecessor)|FLAG );
                         // Now owner of predecessor is waiting for _us_ to release its lock
                         release_internal_lock(*predecessor);
                     }
@@ -338,13 +365,13 @@ struct queuing_rw_mutex_impl {
                 tricky_pointer::store(predecessor->my_next, nullptr, std::memory_order_release);
 
                 d1::queuing_rw_mutex::scoped_lock* expected = &s;
-                if( !tricky_pointer::load(s.my_next, std::memory_order_relaxed) && !s.my_mutex->q_tail.compare_exchange_strong(expected, predecessor, std::memory_order_release) ) {
+                if( !tricky_pointer::load(s.my_next, std::memory_order_acquire) && !s.my_mutex->q_tail.compare_exchange_strong(expected, predecessor, std::memory_order_release) ) {
                     spin_wait_while_eq( s.my_next, 0U, std::memory_order_acquire );
                 }
                 __TBB_ASSERT( !(s.my_next.load(std::memory_order_relaxed) & FLAG), "use of corrupted pointer" );
 
-                // ensure acquire semantics of reading 'my_next'
-                if(d1::queuing_rw_mutex::scoped_lock *const l_next = tricky_pointer::load(s.my_next, std::memory_order_acquire) ) { // I->next != nil, TODO: rename to next after clearing up and adapting the n in the comment two lines below
+                // my_next is acquired either with load or spin_wait.
+                if(d1::queuing_rw_mutex::scoped_lock *const l_next = tricky_pointer::load(s.my_next, std::memory_order_relaxed) ) { // I->next != nil, TODO: rename to next after clearing up and adapting the n in the comment two lines below
                     // Equivalent to I->next->prev = I->prev but protected against (prev[n]&FLAG)!=0
                     tmp = tricky_pointer::exchange(l_next->my_prev, predecessor, std::memory_order_release);
                     // I->prev->next = I->next;
@@ -359,9 +386,12 @@ struct queuing_rw_mutex_impl {
                 d1::queuing_rw_mutex::scoped_lock* next = tricky_pointer::load(s.my_next, std::memory_order_acquire);
                 if( !next ) {
                     d1::queuing_rw_mutex::scoped_lock* expected = &s;
-                    if( !s.my_mutex->q_tail.compare_exchange_strong(expected, nullptr, std::memory_order_release) ) {
-                        spin_wait_while_eq( s.my_next, 0U );
-                        next = tricky_pointer::load(s.my_next, std::memory_order_relaxed);
+                    // Release mutex on success otherwise wait for successor publication
+                    if( !s.my_mutex->q_tail.compare_exchange_strong(expected, nullptr,
+                        std::memory_order_release, std::memory_order_relaxed) )
+                    {
+                        spin_wait_while_eq( s.my_next, 0U, std::memory_order_relaxed );
+                        next = tricky_pointer::load(s.my_next, std::memory_order_acquire);
                     } else {
                         goto unlock_self;
                     }
@@ -375,7 +405,8 @@ struct queuing_rw_mutex_impl {
             unblock_or_wait_on_internal_lock(s, get_flag(tmp));
         }
     done:
-        spin_wait_while_eq( s.my_going, 2U );
+        // Lifetime synchronization, no need to build happens-before relation
+        spin_wait_while_eq( s.my_going, 2U, std::memory_order_relaxed );
 
         s.initialize();
     }
@@ -384,26 +415,41 @@ struct queuing_rw_mutex_impl {
         if ( s.my_state.load(std::memory_order_relaxed) == STATE_ACTIVEREADER ) return true; // Already a reader
 
         ITT_NOTIFY(sync_releasing, s.my_mutex);
-        s.my_state.store(STATE_READER, std::memory_order_relaxed);
-        if( ! tricky_pointer::load(s.my_next, std::memory_order_relaxed)) {
+        d1::queuing_rw_mutex::scoped_lock* next = tricky_pointer::load(s.my_next, std::memory_order_acquire);
+        if( !next ) {
+            s.my_state.store(STATE_READER, std::memory_order_seq_cst);
             // the following load of q_tail must not be reordered with setting STATE_READER above
-            if( &s==s.my_mutex->q_tail.load() ) {
+            if( &s == s.my_mutex->q_tail.load(std::memory_order_seq_cst) ) {
                 unsigned char old_state = STATE_READER;
-                s.my_state.compare_exchange_strong(old_state, STATE_ACTIVEREADER, std::memory_order_release);
+                // When this reader is signaled by previous actor it acquires the mutex.
+                // We need to build happens-before relation with all other coming readers that will read our ACTIVEREADER
+                // without blocking on my_going. Therefore, we need to publish ACTIVEREADER with release semantics.
+                // On fail it is relaxed, because we will build happens-before on my_going.
+                s.my_state.compare_exchange_strong(old_state, STATE_ACTIVEREADER, std::memory_order_release, std::memory_order_relaxed);
                 if( old_state==STATE_READER )
                     return true; // Downgrade completed
             }
             /* wait for the next to register */
-            spin_wait_while_eq( s.my_next, 0U );
+            spin_wait_while_eq(s.my_next, 0U, std::memory_order_relaxed);
+            next = tricky_pointer::load(s.my_next, std::memory_order_acquire);
         }
-        d1::queuing_rw_mutex::scoped_lock *const next = tricky_pointer::load(s.my_next, std::memory_order_acquire);
+
         __TBB_ASSERT( next, "still no successor at this point!" );
-        if( next->my_state & STATE_COMBINED_WAITINGREADER )
+        if( next->my_state.load(std::memory_order_relaxed) & STATE_COMBINED_WAITINGREADER )
             next->my_going.store(1U, std::memory_order_release);
-        else if( next->my_state==STATE_UPGRADE_WAITING )
+        // If the next is STATE_UPGRADE_WAITING, it is expected to acquire all other released readers via release
+        // sequence in next->my_state. In that case, we need to preserve release sequence in next->my_state
+        // contributed by other reader. So, there are two approaches not to break the release sequence:
+        //   1. Use read-modify-write (exchange) operation to store with release the UPGRADE_LOSER state;
+        //   2. Acquire the release sequence and store the sequence and UPGRADE_LOSER state.
+        // The second approach seems better on x86 because it does not involve interlocked operations.
+        // Therefore, we read next->my_state with acquire while it is not required for else branch to get the 
+        // release sequence.
+        else if( next->my_state.load(std::memory_order_acquire)==STATE_UPGRADE_WAITING )
             // the next waiting for upgrade means this writer was upgraded before.
-            next->my_state.store(STATE_UPGRADE_LOSER, std::memory_order_relaxed);
-        s.my_state.store(STATE_ACTIVEREADER, std::memory_order_relaxed);
+            // To safe release sequence on next->my_state read it with acquire
+            next->my_state.store(STATE_UPGRADE_LOSER, std::memory_order_release);
+        s.my_state.store(STATE_ACTIVEREADER, std::memory_order_release);
         return true;
     }
 
@@ -419,16 +465,20 @@ struct queuing_rw_mutex_impl {
         queuing_rw_mutex::scoped_lock* me = &s;
 
         ITT_NOTIFY(sync_releasing, s.my_mutex);
-        s.my_state.store(STATE_UPGRADE_REQUESTED, std::memory_order_relaxed);
+        // Publish ourselves into my_state that other UPGRADE_WAITING actors can acquire our state.
+        s.my_state.store(STATE_UPGRADE_REQUESTED, std::memory_order_release);
     requested:
         __TBB_ASSERT( !(s.my_next.load(std::memory_order_relaxed) & FLAG), "use of corrupted pointer!" );
         acquire_internal_lock(s);
         d1::queuing_rw_mutex::scoped_lock* expected = &s;
-        if( !s.my_mutex->q_tail.compare_exchange_strong(expected, tricky_pointer(me)|FLAG, std::memory_order_release) ) {
-            spin_wait_while_eq( s.my_next, 0U );
+        if( !s.my_mutex->q_tail.compare_exchange_strong(expected, tricky_pointer(me)|FLAG, std::memory_order_acq_rel) ) {
+            spin_wait_while_eq( s.my_next, 0U, std::memory_order_relaxed );
             queuing_rw_mutex::scoped_lock * next;
             next = tricky_pointer::fetch_add(s.my_next, FLAG, std::memory_order_acquire);
-            unsigned short n_state = next->my_state;
+            // While we were READER the next READER might reach STATE_UPGRADE_WAITING state.
+            // Therefore, it did not build happens before relation with us and we need to acquire the 
+            // next->my_state to build the happens before relation ourselves
+            unsigned short n_state = next->my_state.load(std::memory_order_acquire);
             /* the next reader can be blocked by our state. the best thing to do is to unblock it */
             if( n_state & STATE_COMBINED_WAITINGREADER )
                 next->my_going.store(1U, std::memory_order_release);
@@ -439,7 +489,7 @@ struct queuing_rw_mutex_impl {
                 // save next|FLAG for simplicity of following comparisons
                 tmp = tricky_pointer(next)|FLAG;
                 for( atomic_backoff b; tricky_pointer::load(s.my_next, std::memory_order_relaxed)==tmp; b.pause() ) {
-                    if( s.my_state & STATE_COMBINED_UPGRADING ) {
+                    if( s.my_state.load(std::memory_order_acquire) & STATE_COMBINED_UPGRADING ) {
                         if( tricky_pointer::load(s.my_next, std::memory_order_acquire)==tmp )
                             tricky_pointer::store(s.my_next, next, std::memory_order_relaxed);
                         goto waiting;
@@ -458,7 +508,11 @@ struct queuing_rw_mutex_impl {
         } // if( this != my_mutex->q_tail... )
         {
             unsigned char old_state = STATE_UPGRADE_REQUESTED;
-            s.my_state.compare_exchange_strong(old_state, STATE_UPGRADE_WAITING, std::memory_order_acquire);
+            // If we reach STATE_UPGRADE_WAITING state we do not build happens-before relation with READER on
+            // left. We delegate this responsibility to READER on left when it try upgrading. Therefore, we are releasing
+            // on success.
+            // Otherwise, on fail, we already acquired the next->my_state.
+            s.my_state.compare_exchange_strong(old_state, STATE_UPGRADE_WAITING, std::memory_order_release, std::memory_order_relaxed);
         }
     waiting:
         __TBB_ASSERT( !( s.my_next.load(std::memory_order_relaxed) & FLAG ), "use of corrupted pointer!" );
@@ -477,11 +531,14 @@ struct queuing_rw_mutex_impl {
                 // While the predecessor pointer (my_prev) is in use (FLAG is set), we can safely update the node`s state.
                 // Corrupted pointer transitions responsibility to release the predecessor`s node on us.
                 unsigned char old_state = STATE_UPGRADE_REQUESTED;
-                predecessor->my_state.compare_exchange_strong(old_state, STATE_UPGRADE_WAITING, std::memory_order_release);
+                // Try to build happens before with the upgrading READER on left. If fail, the predecessor state is not
+                // important for us because it will acquire our state.
+                predecessor->my_state.compare_exchange_strong(old_state, STATE_UPGRADE_WAITING, std::memory_order_release,
+                    std::memory_order_relaxed);
             }
             if( !success ) {
                 // Responsibility transition, the one who reads uncorrupted my_prev will do release.
-                tmp = tricky_pointer::compare_exchange_strong(s.my_prev, tricky_pointer(predecessor)|FLAG, predecessor, std::memory_order_release);
+                tmp = tricky_pointer::compare_exchange_strong(s.my_prev, tricky_pointer(predecessor)|FLAG, predecessor, std::memory_order_acquire);
                 if( tricky_pointer(tmp) & FLAG ) {
                     tricky_pointer::spin_wait_while_eq(s.my_prev, predecessor);
                     predecessor = tricky_pointer::load(s.my_prev, std::memory_order_relaxed);
@@ -509,9 +566,6 @@ struct queuing_rw_mutex_impl {
         // now wait for the predecessor to finish working with my fields
         spin_wait_while_eq( s.my_going, 2U );
 
-        // Acquire critical section indirectly from previous owner or directly from predecessor (TODO: not clear).
-        atomic_fence(std::memory_order_acquire); // on either "my_mutex->q_tail" or "my_going" (TODO: not clear)
-
         bool result = ( s.my_state != STATE_UPGRADE_LOSER );
         s.my_state.store(STATE_WRITER, std::memory_order_relaxed);
         s.my_going.store(1U, std::memory_order_relaxed);

test/conformance/conformance_mutex.cpp

@@ -40,8 +40,12 @@ TEST_CASE("Basic Locable requirement test") {
     GeneralTest<oneapi::tbb::spin_rw_mutex>("Spin RW Mutex");
     GeneralTest<oneapi::tbb::queuing_mutex>("Queuing Mutex");
     GeneralTest<oneapi::tbb::queuing_rw_mutex>("Queuing RW Mutex");
+    // TODO: Consider adding Thread Sanitizer (note that accesses inside the transaction
+    // considered as races by Thread Sanitizer)
+#if !__TBB_USE_THREAD_SANITIZER
     GeneralTest<oneapi::tbb::speculative_spin_mutex>("Speculative Spin Mutex");
     GeneralTest<oneapi::tbb::speculative_spin_rw_mutex>("Speculative Spin RW Mutex");
+#endif
     // NullMutexes
     GeneralTest<oneapi::tbb::null_mutex, utils::AtomicCounter<oneapi::tbb::null_mutex>>("Null Mutex", false);
     GeneralTest<oneapi::tbb::null_rw_mutex, utils::AtomicCounter<oneapi::tbb::null_rw_mutex>>("Null RW Mutex", false);
@@ -56,8 +60,10 @@ TEST_CASE("Lockable requirement test") {
     TestTryAcquire<oneapi::tbb::spin_rw_mutex>("Spin RW Mutex");
     TestTryAcquire<oneapi::tbb::queuing_mutex>("Queuing Mutex");
     TestTryAcquire<oneapi::tbb::queuing_rw_mutex>("Queuing RW Mutex");
+#if !__TBB_USE_THREAD_SANITIZER
     TestTryAcquire<oneapi::tbb::speculative_spin_mutex>("Speculative Spin Mutex");
     TestTryAcquire<oneapi::tbb::speculative_spin_rw_mutex>("Speculative Spin RW Mutex");
+#endif
     TestTryAcquire<oneapi::tbb::null_mutex>("Null Mutex");
 }
 
@@ -67,16 +73,18 @@ TEST_CASE("Shared mutexes (reader/writer) test") {
     // General reader writer capabilities + upgrade/downgrade
     TestReaderWriterLock<oneapi::tbb::spin_rw_mutex>("Spin RW Mutex");
     TestReaderWriterLock<oneapi::tbb::queuing_rw_mutex>("Queuing RW Mutex");
-    TestReaderWriterLock<oneapi::tbb::speculative_spin_rw_mutex>("Speculative Spin RW Mutex");
     TestNullRWMutex<oneapi::tbb::null_rw_mutex>("Null RW Mutex");
     // Single threaded read/write try_acquire operations
     TestTryAcquireReader<oneapi::tbb::spin_rw_mutex>("Spin RW Mutex");
     TestTryAcquireReader<oneapi::tbb::queuing_rw_mutex>("Queuing RW Mutex");
     TestRWStateMultipleChange<oneapi::tbb::spin_rw_mutex>("Spin RW Mutex");
     TestRWStateMultipleChange<oneapi::tbb::queuing_rw_mutex>("Queuing RW Mutex");
+    TestTryAcquireReader<oneapi::tbb::null_rw_mutex>("Null RW Mutex");
+#if !__TBB_USE_THREAD_SANITIZER
+    TestReaderWriterLock<oneapi::tbb::speculative_spin_rw_mutex>("Speculative Spin RW Mutex");
     TestTryAcquireReader<oneapi::tbb::speculative_spin_rw_mutex>("Speculative Spin RW Mutex");
     TestRWStateMultipleChange<oneapi::tbb::speculative_spin_rw_mutex>("Speculative Spin RW Mutex");
-    TestTryAcquireReader<oneapi::tbb::null_rw_mutex>("Null RW Mutex");
+#endif
 }
 
 //! Testing ISO C++ Mutex and Shared Mutex requirements.

test/tbb/test_join_node.cpp

@@ -247,7 +247,7 @@ void test(ConnectJoinNodeFunc&& connect_join_node) {
 
     remove_edge(bn3, input_port<2>(*jn));
     remove_edge(bn2, input_port<1>(*jn));
-    remove_edge(bn1, input_port<0>(*jn));
+    remove_edge(bn1, *jn); //Removes an edge between a sender and port 0 of a multi-input successor.
     remove_edge(*jn, qn);
 
     run_and_check(g, bn1, bn2, bn3, qn, /*expected=*/false);

test/tbb/test_limiter_node.cpp

@@ -125,6 +125,18 @@ struct put_dec_body : utils::NoAssign {
 
 };
 
+template< typename Sender, typename Receiver >
+void make_edge_impl(Sender& sender, Receiver& receiver){
+#if __GNUC__ < 12 && !TBB_USE_DEBUG
+    // Seemingly, GNU compiler generates incorrect code for the call of limiter.register_successor in release (-03)
+    // The function pointer to make_edge workarounds the issue for unknown reason
+    auto make_edge_ptr = tbb::flow::make_edge<int>;
+    make_edge_ptr(sender, receiver);
+#else
+    tbb::flow::make_edge(sender, receiver);
+#endif
+}
+
 template< typename T >
 void test_puts_with_decrements( int num_threads, tbb::flow::limiter_node< T >& lim , tbb::flow::graph& g) {
     parallel_receiver<T> r(g);
@@ -355,14 +367,7 @@ void test_reserve_release_messages() {
     broad.try_put(1); //failed message retrieved.
     g.wait_for_all();
 
-#if __GNUC__ && __GNUC__ < 12 && !TBB_USE_DEBUG
-    // Seemingly, GNU compiler generates incorrect code for the call of limiter.register_successor in release (-03)
-    // The function pointer to make_edge workarounds the issue for unknown reason
-    auto make_edge_ptr = make_edge<int>;
-    make_edge_ptr(limit, output_queue); //putting the successor back
-#else
-    make_edge(limit, output_queue); //putting the successor back
-#endif
+    make_edge_impl(limit, output_queue); //putting the successor back
 
     broad.try_put(1);  //drop the count
 
@@ -445,7 +450,9 @@ void test_try_put_without_successors() {
     int try_put_num{3};
     tbb::flow::buffer_node<int> bn(g);
     tbb::flow::limiter_node<int> ln(g, try_put_num);
+
     tbb::flow::make_edge(bn, ln);
+
     int i = 1;
     for (; i <= try_put_num; i++)
         bn.try_put(i);
@@ -457,7 +464,9 @@ void test_try_put_without_successors() {
             return int{};
         }
     );
-    tbb::flow::make_edge(ln, fn);
+
+    make_edge_impl(ln, fn);
+
     g.wait_for_all();
     CHECK((counter == i * try_put_num / 2));
 

test/tbb/test_mutex.cpp

@@ -33,7 +33,9 @@
 //! \brief Test for [mutex.spin_mutex mutex.spin_rw_mutex mutex.queuing_mutex mutex.queuing_rw_mutexmutex.speculative_spin_mutex mutex.speculative_spin_rw_mutex] specifications
 
 // TODO: Investigate why RTM doesn't work on some macOS.
-#if __TBB_TSX_INTRINSICS_PRESENT && !__APPLE__
+// TODO: Consider adding Thread Sanitizer (note that accesses inside the transaction
+// considered as races by Thread Sanitizer)
+#if __TBB_TSX_INTRINSICS_PRESENT && !__APPLE__ && !__TBB_USE_THREAD_SANITIZER
 
 inline static bool IsInsideTx() {
     return _xtest() != 0;

test/tbb/test_resumable_tasks.cpp

@@ -255,6 +255,12 @@ private:
 
 // Simple test for basic resumable tasks functionality
 void TestSuspendResume() {
+#if __TBB_USE_SANITIZERS
+    constexpr int iter_size = 100;
+#else
+    constexpr int iter_size = 50000;
+#endif
+
     std::atomic<int> global_epoch; global_epoch = 0;
     EpochAsyncActivity async(4, global_epoch);
 
@@ -263,7 +269,7 @@ void TestSuspendResume() {
     inner_par_iters = outer_par_iters = 0;
 
     tbb::parallel_for(0, N, [&](int) {
-        for (int i = 0; i < 100; ++i) {
+        for (int i = 0; i < iter_size; ++i) {
             ets_fiber.local() = i;
 
             int local_epoch;
@@ -286,7 +292,7 @@ void TestSuspendResume() {
         ++outer_par_iters;
     });
     CHECK(outer_par_iters == N);
-    CHECK(inner_par_iters == N*N*100);
+    CHECK(inner_par_iters == N*N*iter_size);
 }
 
 // During cleanup external thread's local task pool may

test/tbb/test_task_group.cpp

@@ -1173,6 +1173,23 @@ TEST_CASE("Task handle run"){
     CHECK_MESSAGE(h == nullptr, "Delayed task can be executed only once");
 }
 
+//! Basic test for task handle
+//! \brief \ref interface \ref requirement
+TEST_CASE("Task handle run_and_wait"){
+    tbb::task_handle h;
+
+    tbb::task_group tg;
+    bool run {false};
+
+    h = tg.defer([&]{
+        run = true;
+    });
+    CHECK_MESSAGE(run == false, "delayed task should not be run until run(task_handle) is called");
+    tg.run_and_wait(std::move(h));
+    CHECK_MESSAGE(run == true, "Delayed task should be completed when task_group::wait exits");
+
+    CHECK_MESSAGE(h == nullptr, "Delayed task can be executed only once");
+}
 //! Test for empty check
 //! \brief \ref interface
 TEST_CASE("Task handle empty check"){
@@ -1308,7 +1325,7 @@ TEST_CASE("task_handle cannot be scheduled into different task_group"){
 //! The test for error in task_handle being scheduled into task_group different from one it was created from
 //! \brief \ref requirement
 TEST_CASE("task_handle cannot be scheduled into other task_group of the same context"
-        * doctest::should_fail()    //Implementation is no there yet, as it is not clear that is the expected behaviour
+        * doctest::should_fail()    //Implementation is no there yet, as it is not clear that is the expected behavior
         * doctest::skip()           //skip the test for now, to not pollute the test log
 )
 {