Re: Futexes ain't fast

On 8/29/2024 7:30 AM, Bonita Montero wrote:

This is by far the fastest code:
 #if defined(_WIN32)
     HANDLE hSem = CreateSemaphoreA( nullptr, 0, 0x7FFFFFFF, nullptr );
     auto acquire = [&] { WaitForSingleObject( hSem, INFINITE ); };
     auto release = [&] { ReleaseSemaphore( hSem, 1, nullptr ); };

^^^^^^^^^^^^^^^^^^^^^^
Actually, release can be SetEvent because the mutex logic only needs a binary semaphore, or an event on windows. However, I cannot remember if an event is "faster" than a semaphore on windows or not. Still do not know how windows handles it's futexes wrt internal impl. WaitOnAddress and such.

#elif defined(__unix__)
     sem_t sem;
     sem_init( &sem, 0, 0 );
     auto acquire = [&] { while( sem_wait( &sem ) == EINTR ); };
     auto release = [&] { sem_post( &sem ); };
#endif
     tests.emplace_back( "CT-Sema", [&]
         {
             if( ctFutex.exchange( 1, memory_order_acquire ) )
                 while( ctFutex.exchange( 2, memory_order_acquire ) )
                     acquire();
             if( ctFutex.exchange( 0, memory_order_release ) == 2 )
                 release();
         } );