SPU/vm: Improve vm::range_lock a bit

Use some prefetching
Use optimistic locking

rpcs3/Emu/Cell/SPUThread.cpp

@@ -1736,6 +1736,8 @@ void spu_thread::push_snr(u32 number, u32 value)
 
 void spu_thread::do_dma_transfer(spu_thread* _this, const spu_mfc_cmd& args, u8* ls)
 {
+	perf_meter<"DMA"_u32> perf_;
+
 	const bool is_get = (args.cmd & ~(MFC_BARRIER_MASK | MFC_FENCE_MASK | MFC_START_MASK)) == MFC_GET_CMD;
 
 	u32 eal = args.eal;
@@ -1834,14 +1836,8 @@ void spu_thread::do_dma_transfer(spu_thread* _this, const spu_mfc_cmd& args, u8*
 
 	if ((!g_use_rtm && !is_get) || g_cfg.core.spu_accurate_dma)  [[unlikely]]
 	{
-		perf_meter<"ADMA_GET"_u64> perf_get;
-		perf_meter<"ADMA_PUT"_u64> perf_put = perf_get;
-
-		if (!g_cfg.core.spu_accurate_dma) [[likely]]
-		{
-			perf_put.reset();
-			perf_get.reset();
-		}
+		perf_meter<"ADMA_GET"_u64> perf_get = perf_;
+		perf_meter<"ADMA_PUT"_u64> perf_put = perf_;
 
 		cpu_thread* _cpu = _this ? _this : get_current_cpu_thread();
 
@@ -1864,6 +1860,8 @@ void spu_thread::do_dma_transfer(spu_thread* _this, const spu_mfc_cmd& args, u8*
 			range_lock = _this->range_lock;
 		}
 
+		_m_prefetchw(range_lock);
+
 		for (u32 size = args.size, size0; is_get; size -= size0, dst += size0, src += size0, eal += size0)
 		{
 			size0 = std::min<u32>(128 - (eal & 127), std::min<u32>(size, 128));
@@ -2161,8 +2159,13 @@ void spu_thread::do_dma_transfer(spu_thread* _this, const spu_mfc_cmd& args, u8*
 			//std::atomic_thread_fence(std::memory_order_seq_cst);
 			return;
 		}
+		else
+		{
+			perf_put.reset();
+			perf_get.reset();
+		}
 
-		perf_meter<"DMA_PUT"_u64> perf2;
+		perf_meter<"DMA_PUT"_u64> perf2 = perf_;
 
 		switch (u32 size = args.size)
 		{

rpcs3/Emu/Memory/vm.cpp

@@ -167,6 +167,8 @@ namespace vm
 
 		for (u64 i = 0;; i++)
 		{
+			range_lock->store(begin | (u64{size} << 32));
+
 			const u64 lock_val = g_range_lock.load();
 			const u64 is_share = g_shmem[begin >> 16].load();
 
@@ -188,18 +190,18 @@ namespace vm
 
 			if (addr + size <= lock_addr || addr >= lock_addr + lock_size) [[likely]]
 			{
-				range_lock->store(begin | (u64{size} << 32));
-
 				const u64 new_lock_val = g_range_lock.load();
 
 				if (!new_lock_val || new_lock_val == lock_val) [[likely]]
 				{
 					break;
 				}
-
-				range_lock->release(0);
 			}
 
+			// Wait a bit before accessing g_mutex
+			range_lock->store(0);
+			busy_wait(200);
+
 			std::shared_lock lock(g_mutex, std::try_to_lock);
 
 			if (!lock && i < 15)

rpcs3/Emu/Memory/vm_locking.h

@@ -41,13 +41,20 @@ namespace vm
 
 	void range_lock_internal(atomic_t<u64, 64>* range_lock, u32 begin, u32 size);
 
-	// Lock memory range
+	// Lock memory range ignoring memory protection (Size!=0 also implies aligned begin)
 	template <uint Size = 0>
 	FORCE_INLINE void range_lock(atomic_t<u64, 64>* range_lock, u32 begin, u32 _size)
 	{
+		// Optimistic locking.
+		// Note that we store the range we will be accessing, without any clamping.
+		range_lock->store(begin | (u64{_size} << 32));
+
+		// Old-style conditional constexpr
 		const u32 size = Size ? Size : _size;
+
 		const u64 lock_val = g_range_lock.load();
 		const u64 is_share = g_shmem[begin >> 16].load();
+
 		#ifndef _MSC_VER
 		__asm__(""); // Tiny barrier
 		#endif
@@ -59,7 +66,7 @@ namespace vm
 
 		// Optimization: if range_locked is not used, the addr check will always pass
 		// Otherwise, g_shmem is unchanged and its value is reliable to read
-		if ((lock_val >> range_pos) == (range_locked >> range_pos)) [[likely]]
+		if ((lock_val >> range_pos) == (range_locked >> range_pos))
 		{
 			lock_size = 128;
 
@@ -72,19 +79,15 @@ namespace vm
 
 		if (addr + size <= lock_addr || addr >= lock_addr + lock_size) [[likely]]
 		{
-			// Optimistic locking.
-			// Note that we store the range we will be accessing, without any clamping.
-			range_lock->store(begin | (u64{size} << 32));
-
 			const u64 new_lock_val = g_range_lock.load();
 
 			if (!new_lock_val || new_lock_val == lock_val) [[likely]]
 			{
 				return;
 			}
+		}
 
 		range_lock->release(0);
-		}
 
 		// Fallback to slow path
 		range_lock_internal(range_lock, begin, size);