softgpu: Use persistent bin task state.

It's constant, so it's better to avoid the copying and allocation.  A
small win, but removes new from the profile.

Common/Thread/ThreadManager.cpp

@@ -105,7 +105,7 @@ bool ThreadManager::TeardownTask(Task *task, bool enqueue) {
 
 	if (task->Cancellable()) {
 		task->Cancel();
-		delete task;
+		task->Release();
 		return true;
 	}
 
@@ -174,7 +174,7 @@ static void WorkerThreadFunc(GlobalThreadContext *global, ThreadContext *thread)
 		// responsibility of the ThreadManager (although it could be!).
 		if (task) {
 			task->Run();
-			delete task;
+			task->Release();
 
 			// Reduce the queue size once complete.
 			thread->queue_size--;

Common/Thread/ThreadManager.h

@@ -19,6 +19,7 @@ public:
 	virtual bool Cancellable() { return false; }
 	virtual void Cancel() {}
 	virtual uint64_t id() { return 0; }
+	virtual void Release() { delete this; }
 };
 
 class Waitable {

GPU/Software/BinManager.cpp

@@ -105,6 +105,10 @@ public:
 		notify_->Drain();
 	}
 
+	void Release() override {
+		// Don't delete, this is statically allocated.
+	}
+
 private:
 	void ProcessItems() {
 		while (!items_.Empty()) {
@@ -131,8 +135,11 @@ BinManager::BinManager() {
 		s = false;
 
 	int maxInitTasks = std::min(g_threadManager.GetNumLooperThreads(), MAX_POSSIBLE_TASKS);
-	for (int i = 0; i < maxInitTasks; ++i)
+	for (int i = 0; i < maxInitTasks; ++i) {
 		taskQueues_[i].Setup();
+		for (DrawBinItemsTask *&task : taskLists_[i].tasks)
+			task = new DrawBinItemsTask(waitable_, taskQueues_[i], taskStatus_[i], states_);
+	}
 	states_.Setup();
 	cluts_.Setup();
 	queue_.Setup();
@@ -140,6 +147,11 @@ BinManager::BinManager() {
 
 BinManager::~BinManager() {
 	delete waitable_;
+
+	for (int i = 0; i < MAX_POSSIBLE_TASKS; ++i) {
+		for (DrawBinItemsTask *task : taskLists_[i].tasks)
+			delete task;
+	}
 }
 
 void BinManager::UpdateState() {
@@ -335,8 +347,7 @@ void BinManager::Drain() {
 
 			waitable_->Fill();
 			taskStatus_[i] = true;
-			DrawBinItemsTask *task = new DrawBinItemsTask(waitable_, taskQueues_[i], taskStatus_[i], states_);
-			g_threadManager.EnqueueTaskOnThread(i, task, true);
+			g_threadManager.EnqueueTaskOnThread(i, taskLists_[i].Next(), true);
 			enqueues_++;
 		}
 

GPU/Software/BinManager.h

@@ -152,6 +152,19 @@ union BinClut {
 	uint8_t readable[1024];
 };
 
+struct BinTaskList {
+	// We shouldn't ever need more than two at once, since we use an atomic to run one at a time.
+	// A second could run due to overlap during teardown.
+	static constexpr int N = 2;
+
+	DrawBinItemsTask *tasks[N]{};
+	int count = 0;
+
+	DrawBinItemsTask *Next() {
+		return tasks[count % N];
+	}
+};
+
 class BinManager {
 public:
 	BinManager();
@@ -204,6 +217,7 @@ private:
 	bool tasksSplit_ = false;
 	std::vector<BinCoords> taskRanges_;
 	BinItemQueue taskQueues_[MAX_POSSIBLE_TASKS];
+	BinTaskList taskLists_[MAX_POSSIBLE_TASKS];
 	std::atomic<bool> taskStatus_[MAX_POSSIBLE_TASKS];
 	BinWaitable *waitable_ = nullptr;