From 32767ce7f97305b4c925b8279c5696af1ef1abb1 Mon Sep 17 00:00:00 2001
From: Eladash <18193363+elad335@users.noreply.github.com>
Date: Mon, 26 Feb 2024 23:38:12 +0200
Subject: [PATCH] SPU LLVM: Improve runtime SPU compilation preferences

---
 Utilities/lockless.h                   | 14 ++++++--
 rpcs3/Emu/Cell/SPUCommonRecompiler.cpp | 50 ++++++++++++++++++++++++--
 2 files changed, 60 insertions(+), 4 deletions(-)

diff --git a/Utilities/lockless.h b/Utilities/lockless.h
index f49544ea3d4c..e361434a7fa1 100644
--- a/Utilities/lockless.h
+++ b/Utilities/lockless.h
@@ -379,8 +379,8 @@ class lf_queue final
 		return m_head != 0;
 	}
 
-	template <typename... Args>
-	void push(Args&&... args)
+	template <bool Notify = true, typename... Args>
+	bool push(Args&&... args)
 	{
 		auto oldv = m_head.load();
 		auto item = new lf_queue_item<T>(load(oldv), std::forward<Args>(args)...);
@@ -393,6 +393,16 @@ class lf_queue final
 		if (!oldv)
 		{
 			// Notify only if queue was empty
+			notify(true);
+		}
+
+		return !!oldv;
+	}
+
+	void notify(bool force = false)
+	{
+		if (force || operator bool())
+		{
 			utils::bless<atomic_t<u32>>(&m_head)[1].notify_one();
 		}
 	}
diff --git a/rpcs3/Emu/Cell/SPUCommonRecompiler.cpp b/rpcs3/Emu/Cell/SPUCommonRecompiler.cpp
index 7b9cba2347f6..41f80e4e77ae 100644
--- a/rpcs3/Emu/Cell/SPUCommonRecompiler.cpp
+++ b/rpcs3/Emu/Cell/SPUCommonRecompiler.cpp
@@ -4570,11 +4570,15 @@ struct spu_llvm
 		}
 
 		u32 worker_index = 0;
+		u32 notify_compile_count = 0;
+		std::vector<u8> notify_compile(worker_count);
 
 		m_workers = make_single<named_thread_group<spu_llvm_worker>>("SPUW.", worker_count);
 		auto workers_ptr = m_workers.load();
 		auto& workers = *workers_ptr;
 
+		usz add_count = 65535;
+
 		while (thread_ctrl::state() != thread_state::aborting)
 		{
 			for (const auto& pair : registered.pop_all())
@@ -4585,14 +4589,30 @@ struct spu_llvm
 				const auto lock = prof_mutex.init_always([&]{});
 
 				// Register new blocks to collect samples
-				samples.emplace(pair.first, 0);
+				// If the sample count is the same, prefer the last block that was pushed
+				samples.emplace(pair.first ? pair.first * 65536 + (add_count-- % 65536) : pair.first * 65536, 0);
 			}
 
 			if (enqueued.empty())
 			{
+				// Send pending notifications
+				if (notify_compile_count)
+				{
+					for (usz i = 0; i < worker_count; i++)
+					{
+						if (notify_compile[i])
+						{
+							(workers.begin() + (worker_index % worker_count))->registered.notify();
+						}
+					}
+				}
+
 				// Interrupt profiler thread and put it to sleep
 				static_cast<void>(prof_mutex.reset());
 				thread_ctrl::wait_on(utils::bless<atomic_t<u32>>(&registered)[1], 0);
+				add_count = 65535; // Reset count
+				std::fill(notify_compile.begin(), notify_compile.end(), 0); // Reset notification flags
+				notify_compile_count = 0;
 				continue;
 			}
 
@@ -4620,8 +4640,34 @@ struct spu_llvm
 			// Remove item from the queue
 			enqueued.erase(found_it);
 
+			// Prefer using an inactive thread
+			for (usz i = 0; i < worker_count && !!(workers.begin() + (worker_index % worker_count))->registered; i++)
+			{
+				worker_index++;
+			}
+
 			// Push the workload
-			(workers.begin() + (worker_index++ % worker_count))->registered.push(reinterpret_cast<u64>(_old), &func);
+			const bool notify = (workers.begin() + (worker_index % worker_count))->registered.template push<false>(reinterpret_cast<u64>(_old), &func);
+
+			if (notify && !notify_compile[worker_index % worker_count])
+			{
+				notify_compile[worker_index % worker_count] = 1;
+				notify_compile_cout++;
+
+				if (notify_compile_count == notify_compile.size())
+				{
+					// Notify all
+					for (usz i = 0; i < worker_count; i++)
+					{
+						(workers.begin() + (i % worker_count))->registered.notify();
+					}
+
+					std::fill(notify_compile.begin(), notify_compile.end(), 0); // Reset notification flags
+					notify_compile_count = 0;
+				}
+			}
+
+			worker_index++;
 		}
 
 		static_cast<void>(prof_mutex.init_always([&]{ samples.clear(); }));