Use atomic list for event subscribers

Summary:
Replace the *copy on write* vector with an atomic pointer to a linked list.

This allows to publish without locking a mutex, at the cost of the slower traversal of a linked list (a vector has better locality).

At the moment, the typical use case is to have one subscriber, meaning that the afforementioned slower traversal is not a problem.

Adding subscribers is implemented as atomic *compare and swap.*

Reviewed By: SidharthGuglani

Differential Revision: D15546964

fbshipit-source-id: 41bfa41f1ac6be5c9b6bf4288ea3271ee995877e

Author: davidaurelio

ReactCommon/yoga/yoga/event/event.cpp

@@ -5,51 +5,57 @@
  * file in the root directory of this source tree.
  */
 #include "event.h"
+#include <atomic>
 #include <memory>
 #include <stdexcept>
-#include <mutex>
-
-#include <iostream>
 
 namespace facebook {
 namespace yoga {
 
 namespace {
 
-std::mutex& eventSubscribersMutex() {
-  static std::mutex subscribersMutex;
-  return subscribersMutex;
-}
+struct Node {
+  std::function<Event::Subscriber> subscriber = nullptr;
+  Node* next = nullptr;
+
+  Node(std::function<Event::Subscriber>&& subscriber)
+      : subscriber{std::move(subscriber)} {}
+};
 
-std::shared_ptr<Event::Subscribers>& eventSubscribers() {
-  static auto subscribers = std::make_shared<Event::Subscribers>();
-  return subscribers;
+std::atomic<Node*> subscribers{nullptr};
+
+Node* push(Node* newHead) {
+  Node* oldHead;
+  do {
+    oldHead = subscribers.load(std::memory_order_relaxed);
+    if (newHead != nullptr) {
+      newHead->next = oldHead;
+    }
+  } while (!subscribers.compare_exchange_weak(
+      oldHead, newHead, std::memory_order_release, std::memory_order_relaxed));
+  return oldHead;
 }
 
 } // namespace
 
 void Event::reset() {
-  eventSubscribers() = std::make_shared<Event::Subscribers>();
+  auto head = push(nullptr);
+  while (head != nullptr) {
+    auto current = head;
+    head = head->next;
+    delete current;
+  }
 }
 
 void Event::subscribe(std::function<Subscriber>&& subscriber) {
-  std::lock_guard<std::mutex> guard(eventSubscribersMutex());
-  eventSubscribers() =
-      std::make_shared<Event::Subscribers>(*eventSubscribers());
-  eventSubscribers()->push_back(subscriber);
+  push(new Node{std::move(subscriber)});
 }
 
 void Event::publish(const YGNode& node, Type eventType, const Data& eventData) {
-  std::shared_ptr<Event::Subscribers> subscribers;
-  {
-    std::lock_guard<std::mutex> guard(eventSubscribersMutex());
-    subscribers = eventSubscribers();
-  }
-
-  for (auto& subscriber : *subscribers) {
-    if (subscriber) {
-      subscriber(node, eventType, eventData);
-    }
+  for (auto subscriber = subscribers.load(std::memory_order_relaxed);
+       subscriber != nullptr;
+       subscriber = subscriber->next) {
+    subscriber->subscriber(node, eventType, eventData);
   }
 }