From 4b82006984cabba32ad3e711fdc1a76ae6a41bdb Mon Sep 17 00:00:00 2001
From: Eladash <elad3356p@gmail.com>
Date: Fri, 9 Aug 2019 19:35:42 +0300
Subject: [PATCH] ppu: Improve LWSYNC

Block load<->load reordering as real lwsync.
---
 rpcs3/Emu/Cell/Modules/cellSpurs.cpp    |  2 +-
 rpcs3/Emu/Cell/Modules/cellSync.cpp     | 14 +++++++-------
 rpcs3/Emu/Cell/Modules/sys_lwmutex_.cpp |  4 ++--
 rpcs3/Emu/Cell/PPUTranslator.cpp        |  4 ++--
 4 files changed, 12 insertions(+), 12 deletions(-)

diff --git a/rpcs3/Emu/Cell/Modules/cellSpurs.cpp b/rpcs3/Emu/Cell/Modules/cellSpurs.cpp
index f8aa61441f..d4175a4bb1 100644
--- a/rpcs3/Emu/Cell/Modules/cellSpurs.cpp
+++ b/rpcs3/Emu/Cell/Modules/cellSpurs.cpp
@@ -2577,7 +2577,7 @@ s32 _cellSpursWorkloadFlagReceiver(vm::ptr<CellSpurs> spurs, u32 wid, u32 is_set
 		return CELL_SPURS_POLICY_MODULE_ERROR_STAT;
 	}
 
-	std::atomic_thread_fence(std::memory_order_seq_cst);
+	std::atomic_thread_fence(std::memory_order_acq_rel);
 
 	if (s32 res = spurs->wklFlag.flag.atomic_op([spurs, wid, is_set](be_t<u32>& flag) -> s32
 	{
diff --git a/rpcs3/Emu/Cell/Modules/cellSync.cpp b/rpcs3/Emu/Cell/Modules/cellSync.cpp
index 4dd5f38b4e..0babe4424d 100644
--- a/rpcs3/Emu/Cell/Modules/cellSync.cpp
+++ b/rpcs3/Emu/Cell/Modules/cellSync.cpp
@@ -87,7 +87,7 @@ error_code cellSyncMutexLock(ppu_thread& ppu, vm::ptr<CellSyncMutex> mutex)
 		}
 	}
 
-	std::atomic_thread_fence(std::memory_order_release);
+	std::atomic_thread_fence(std::memory_order_acq_rel);
 	return CELL_OK;
 }
 
@@ -196,7 +196,7 @@ error_code cellSyncBarrierTryNotify(vm::ptr<CellSyncBarrier> barrier)
 		return CELL_SYNC_ERROR_ALIGN;
 	}
 
-	std::atomic_thread_fence(std::memory_order_release);
+	std::atomic_thread_fence(std::memory_order_acq_rel);
 
 	if (!barrier->ctrl.atomic_op<&CellSyncBarrier::try_notify>())
 	{
@@ -220,7 +220,7 @@ error_code cellSyncBarrierWait(ppu_thread& ppu, vm::ptr<CellSyncBarrier> barrier
 		return CELL_SYNC_ERROR_ALIGN;
 	}
 
-	std::atomic_thread_fence(std::memory_order_release);
+	std::atomic_thread_fence(std::memory_order_acq_rel);
 
 	while (!barrier->ctrl.atomic_op<&CellSyncBarrier::try_wait>())
 	{
@@ -247,7 +247,7 @@ error_code cellSyncBarrierTryWait(vm::ptr<CellSyncBarrier> barrier)
 		return CELL_SYNC_ERROR_ALIGN;
 	}
 
-	std::atomic_thread_fence(std::memory_order_release);
+	std::atomic_thread_fence(std::memory_order_acq_rel);
 
 	if (!barrier->ctrl.atomic_op<&CellSyncBarrier::try_wait>())
 	{
@@ -281,7 +281,7 @@ error_code cellSyncRwmInitialize(vm::ptr<CellSyncRwm> rwm, vm::ptr<void> buffer,
 	rwm->size = buffer_size;
 	rwm->buffer = buffer;
 
-	std::atomic_thread_fence(std::memory_order_release);
+	std::atomic_thread_fence(std::memory_order_acq_rel);
 
 	return CELL_OK;
 }
@@ -453,7 +453,7 @@ error_code cellSyncQueueInitialize(vm::ptr<CellSyncQueue> queue, vm::ptr<u8> buf
 	queue->depth = depth;
 	queue->buffer = buffer;
 
-	std::atomic_thread_fence(std::memory_order_release);
+	std::atomic_thread_fence(std::memory_order_acq_rel);
 
 	return CELL_OK;
 }
@@ -866,7 +866,7 @@ error_code cellSyncLFQueueInitialize(vm::ptr<CellSyncLFQueue> queue, vm::cptr<vo
 			}
 		}
 
-		std::atomic_thread_fence(std::memory_order_release);
+		std::atomic_thread_fence(std::memory_order_acq_rel);
 	}
 	else
 	{
diff --git a/rpcs3/Emu/Cell/Modules/sys_lwmutex_.cpp b/rpcs3/Emu/Cell/Modules/sys_lwmutex_.cpp
index 8a8999f788..61cd194295 100644
--- a/rpcs3/Emu/Cell/Modules/sys_lwmutex_.cpp
+++ b/rpcs3/Emu/Cell/Modules/sys_lwmutex_.cpp
@@ -130,7 +130,7 @@ error_code sys_lwmutex_lock(ppu_thread& ppu, vm::ptr<sys_lwmutex_t> lwmutex, u64
 
 		// recursive locking succeeded
 		lwmutex->recursive_count++;
-		std::atomic_thread_fence(std::memory_order_release);
+		std::atomic_thread_fence(std::memory_order_acq_rel);
 
 		return CELL_OK;
 	}
@@ -290,7 +290,7 @@ error_code sys_lwmutex_trylock(ppu_thread& ppu, vm::ptr<sys_lwmutex_t> lwmutex)
 
 		// recursive locking succeeded
 		lwmutex->recursive_count++;
-		std::atomic_thread_fence(std::memory_order_release);
+		std::atomic_thread_fence(std::memory_order_acq_rel);
 
 		return CELL_OK;
 	}
diff --git a/rpcs3/Emu/Cell/PPUTranslator.cpp b/rpcs3/Emu/Cell/PPUTranslator.cpp
index 30d360a47a..6197a885f1 100644
--- a/rpcs3/Emu/Cell/PPUTranslator.cpp
+++ b/rpcs3/Emu/Cell/PPUTranslator.cpp
@@ -3106,8 +3106,8 @@ void PPUTranslator::LFSUX(ppu_opcode_t op)
 void PPUTranslator::SYNC(ppu_opcode_t op)
 {
 	// sync: Full seq cst barrier
-	// lwsync: Release barrier
-	m_ir->CreateFence(op.l10 ? AtomicOrdering::Release : AtomicOrdering::SequentiallyConsistent);
+	// lwsync: Acq/Release barrier
+	m_ir->CreateFence(op.l10 ? AtomicOrdering::AcquireRelease : AtomicOrdering::SequentiallyConsistent);
 }
 
 void PPUTranslator::LFDX(ppu_opcode_t op)