This typedefs to std::atomic<> when building with true threading, and uses a dummy implementation without.
This lets us use the full range of atomic operations offered by C++11, including explicit specification of memory fences. Using barriers lets the compiler generate more optimal code since currently we are using the quite strict sequential-consistent memory ordering for all operations. ReferenceCount has been changed to use the correct barriers (I hope). This may especially make a difference on weak ordering systems such as ARM.
Over time we should gradually replace the use of AtomicAdjust with the new patomic file.
This came up in #330; the Character destructor caused something to call lock() on a weak pointer to that character, which would induce a ref() and unref() pair, but since the refcount was 0, this would call the destructor and thereby create infinite recursion.
I considered instead calling mark_deleted() inside unref() so that the callbacks get to run before the object is actually deleted, and was_deleted() will become true as soon as unref() reaches 0. However, this would require grabbing the lock in unref() to be fully thread-safe, since we would need to bring the refcount to 0 and mark the object as deleted in one atomic operation, so this would be an unacceptable general performance penalty.
Instead, WeakPointerTo::lock() now atomically increments the reference count if it is not already zero, and returns null otherwise. This should be safe because the object cannot be deleted while the WeakReferenceList lock is held.
To access a WeakPointerTo in a thread-safe way, use something like this:
if (auto ptr = weak_ptr.lock()) {
..use ptr as regular PointerTo
}
The new implementation no longer needs a reference to be stored to all weak pointers on the WeakReferenceList; a mere count of weak pointers is sufficient. Therefore, callbacks theoretically no longer require a WeakPointerTo to be constructed.
The WeakPointerTo class is not actually atomic; it could be made so, but I don't believe it's worth it at this time.
* Allow arbitrary selection of optional keyword arguments in some cases
* Reduce code bloat of generated bindings
* Work around awkward resolution of set_shader_input overloads
* Document the code a bit better
* Make coercion a bit cleaner for reference counted types
* A few optimization tweaks
* Use generic errors in NDEBUG build to reduce size of string pool
rdb
Sam Edwards
tobspr
David Rose
Dave Schuyler
cxgeorge