Fix for D25504 - segfault because of double free()-ing in task deps shutdown code.

[jlpeyton]

Summary

Moves the freeing of the implicit task to above the freeing of all fast memory to prevent the double-free issue.

Fixes: https://reviews.llvm.org/D25504

[pawosm01]

I tested this change for few hours running superlu_taskdep and strassen_taskdep on different number of cores. Didn't fail, so we can assume it solves the problem.

BTW, don't you think that the task->td_dephash pointer should be NULLified after calling __kmp_dephash_free() on it in kmp_taskdeps.cpp line 380?

[AndreyChurbanov]

BTW, don't you think that the task->td_dephash pointer should be NULLified after calling __kmp_dephash_free() on it in kmp_taskdeps.cpp line 380?

I agree. Keeping stale pointer is not good generally, regardless that it is not used. So, writing NULL should be safer. Of cause this cleanup can be done via a separate review.

[jlpeyton]

Committed setting task->td_dephash = NULL without review: r287552

openmp/trunk/runtime/src/kmp_runtime.c

if ( ! is_root ) {
             KMP_DEBUG_ASSERT( TCR_4(__kmp_thread_pool_active_nth) >= 0 );
         }
 
         // Decrement # of [worker] threads in the pool.
         KMP_DEBUG_ASSERT( __kmp_thread_pool_nth > 0 );
         --__kmp_thread_pool_nth;
     }; // if
 
+    __kmp_free_implicit_task(thread);
+
     // Free the fast memory for tasking
     #if USE_FAST_MEMORY
         __kmp_free_fast_memory( thread );
     #endif /* USE_FAST_MEMORY */
 
     __kmp_suspend_uninitialize_thread( thread );
 
     KMP_DEBUG_ASSERT( __kmp_threads[ gtid ] == thread );

 
 #if KMP_AFFINITY_SUPPORTED
     if ( thread->th.th_affin_mask != NULL ) {
         KMP_CPU_FREE( thread->th.th_affin_mask );
         thread->th.th_affin_mask = NULL;
     }; // if
 #endif /* KMP_AFFINITY_SUPPORTED */
 
-    __kmp_free_implicit_task(thread);
     __kmp_reap_team( thread->th.th_serial_team );
     thread->th.th_serial_team = NULL;
     __kmp_free( thread );
 
     KMP_MB();
 
 } // __kmp_reap_thread