Data and Data Copy object pool LIFOs underused

[omor1]

Describe the bug

In data.c, when allocating new parsec_data_t * and parsec_data_copy_t * instances (via parsec_data_new and parsec_data_copy_new, respectively), LIFOs are used as a sort of object pool. These LIFOs are parsec_data_lifo and parsec_data_copies_lifo.

The pseudocode for the new functions is:

object_t* new(void)
{
    object_t* item = (object_t*)parsec_lifo_pop(&object_lifo);
    if( NULL == item) {
        item = PARSEC_OBJ_NEW(object_t);
    } else {
        PARSEC_OBJ_CONSTRUCT(item, object_t);
    }
    return item;
}

The problem is that nothing is ever pushed to parsec_data_lifo and parsec_data_copies_lifo; in fact, there is no use of parsec_data_copies_lifo outside of its initialization and deinitialization and in parsec_data_copy_new. parsec_data_delete does push to parsec_data_lifo, but I have identified no uses of this function in the codebase; parsec_data_destroy is used instead, which ignores parsec_data_lifo.

In essence, these LIFOs are not being used for their intended purpose. They should either be removed if they are not necessary or their functionality restored.

Additional context

I was investigating issues where calls to remote_dep_copy_allocate (which calls parsec_arena_get_copy and whence parsec_data_new and parsec_data_copy_new) takes an inordinate amount of time, 100ms or more. My current hypothesis is high contention on shared data structures—of which these LIFOs are one such possibility.