[shader-decoder] fix RDNA2 shift masking and sprite draws

.github/workflows/workflow.yml

@@ -9,11 +9,17 @@ on:
       - "**"
     paths-ignore:
       - "**/*.md"
+      - "**/*.png"
+      - "**/*.jpeg"
+      - "**/*.jpg"
   pull_request:
     branches:
       - main
     paths-ignore:
       - "**/*.md"
+      - "**/*.png"
+      - "**/*.jpeg"
+      - "**/*.jpg"
   workflow_dispatch:
 
 permissions:

src/SharpEmu.Core/Cpu/Native/DirectExecutionBackend.Imports.cs

@@ -482,6 +482,8 @@ public sealed partial class DirectExecutionBackend
 		catch (Exception ex)
 		{
 			LastError = $"HLE dispatch error for {importStubEntry.Nid}: {ex.GetType().Name}: {ex.Message}";
+			Console.Error.WriteLine($"[LOADER][ERROR] {LastError}");
+			Console.Error.WriteLine($"[LOADER][ERROR] {ex.StackTrace}");
 			cpuContext[CpuRegister.Rax] = 18446744071562199298uL;
 			return 18446744071562199298uL;
 		}
@@ -527,15 +529,14 @@ public sealed partial class DirectExecutionBackend
 		if (dispatchIndex % 100000 == 0)
 		{
 			Console.Error.WriteLine(
-				$"[LOADER][TRACE] Import#{dispatchIndex}: {export.LibraryName}:{export.Name} ({importStubEntry.Nid})");
+				$"[LOADER][TRACE] Import#{dispatchIndex}: {export.LibraryName}:{export.Name} ({importStubEntry.Nid}) " +
+				$"rdi=0x{arg0:X16} rsi=0x{cpuContext[CpuRegister.Rsi]:X16} " +
+				$"rdx=0x{cpuContext[CpuRegister.Rdx]:X16} rcx=0x{cpuContext[CpuRegister.Rcx]:X16} " +
+				$"ret=0x{returnRip:X16}");
 		}
 
-		var previousImportCallFrame = GuestThreadExecution.EnterImportCallFrame(
-			returnRip,
-			(ulong)argPackPtr + 104uL,
-			ActiveGuestReturnSlotAddress);
 		int returnValue;
-		try
+		if (IsNoBlockLeafImport(importStubEntry.Nid))
 		{
 			cpuContext.ClearRaxWriteFlag();
 			returnValue = export.Function(cpuContext);
@@ -544,9 +545,25 @@ public sealed partial class DirectExecutionBackend
 				cpuContext[CpuRegister.Rax] = unchecked((ulong)returnValue);
 			}
 		}
-		finally
+		else
 		{
-			GuestThreadExecution.RestoreImportCallFrame(previousImportCallFrame);
+			var previousImportCallFrame = GuestThreadExecution.EnterImportCallFrame(
+				returnRip,
+				(ulong)argPackPtr + 104uL,
+				ActiveGuestReturnSlotAddress);
+			try
+			{
+				cpuContext.ClearRaxWriteFlag();
+				returnValue = export.Function(cpuContext);
+				if (!cpuContext.WasRaxWritten)
+				{
+					cpuContext[CpuRegister.Rax] = unchecked((ulong)returnValue);
+				}
+			}
+			finally
+			{
+				GuestThreadExecution.RestoreImportCallFrame(previousImportCallFrame);
+			}
 		}
 
 		if (returnValue != (int)OrbisGen2Result.ORBIS_GEN2_OK)
@@ -591,6 +608,28 @@ public sealed partial class DirectExecutionBackend
 		return true;
 	}
 
+	private static bool IsNoBlockLeafImport(string nid) =>
+		nid is
+			"8aI7R7WaOlc" or // sceAmprCommandBufferConstructor
+			"a8uLzYY--tM" or // sceAmprAprCommandBufferConstructor
+			"Qs1xtplKo0U" or // sceAmprAprCommandBufferDestructor
+			"GuchCTefuZw" or // sceAmprCommandBufferDestructor
+			"N-FSPA4S3nI" or // sceAmprCommandBufferSetBuffer
+			"baQO9ez2gL4" or // sceAmprCommandBufferReset
+			"ULvXMDz56po" or // sceAmprCommandBufferClearBuffer
+			"mQ16-QdKv7k" or // sceAmprAprCommandBufferReadFile
+			"vWU-odnS+fU" or // sceAmprMeasureCommandSizeReadFile
+			"sSAUCCU1dv4" or // sceAmprMeasureCommandSizeWriteKernelEventQueue_04_00
+			"C+IEj+BsAFM" or // sceAmprMeasureCommandSizeWriteAddressOnCompletion
+			"tZDDEo2tE5k" or // sceAmprCommandBufferGetSize
+			"GnxKOHEawhk" or // sceAmprCommandBufferGetCurrentOffset
+			"H896Pt-yB4I" or // sceAmprCommandBufferWriteKernelEventQueue_04_00
+			"sJXyWHjP-F8" or // sceAmprCommandBufferWriteAddressOnCompletion
+			"ASoW5WE-UPo" or // sceKernelAprSubmitCommandBufferAndGetResult
+			"rqwFKI4PAiM" or // sceKernelAprWaitCommandBuffer
+			"eE4Szl8sil8" or // sceKernelAprSubmitCommandBuffer
+			"qvMUCyyaCSI";   // sceKernelAprSubmitCommandBufferAndGetId
+
 	private bool ShouldLogImportResult(string nid, OrbisGen2Result result)
 	{
 		var expectedFileProbeMiss =
@@ -599,7 +638,10 @@ public sealed partial class DirectExecutionBackend
 		var expectedTimedWaitTimeout =
 			string.Equals(nid, "27bAgiJmOh0", StringComparison.Ordinal) &&
 			unchecked((int)result) == 60;
-		if (!expectedFileProbeMiss && !expectedTimedWaitTimeout)
+		var expectedMutexTrylockBusy =
+			string.Equals(nid, "K-jXhbt2gn4", StringComparison.Ordinal) &&
+			result == OrbisGen2Result.ORBIS_GEN2_ERROR_BUSY;
+		if (!expectedFileProbeMiss && !expectedTimedWaitTimeout && !expectedMutexTrylockBusy)
 		{
 			return true;
 		}
@@ -635,6 +677,31 @@ public sealed partial class DirectExecutionBackend
 			"7H0iTOciTLo" or
 			"2Z+PpY6CaJg" or
 			"8aI7R7WaOlc" or
+			"zgXifHT9ErY" or // sceVideoOutIsFlipPending
+			"V++UgBtQhn0" or // sceAgcGetDataPacketPayloadAddress
+			"qj7QZpgr9Uw" or // Gen5 graphics type-2 packet
+			"LtTouSCZjHM" or // sceAgcCbNop
+			"k3GhuSNmBLU" or // sceAgcCbDispatch
+			"UZbQjYAwwXM" or // sceAgcCbSetShRegistersDirect
+			"JrtiDtKeS38" or // sceAgcAcbResetQueue
+			"cFazmnXpJOE" or // sceAgcAcbEventWrite
+			"KT-hTp-Ch14" or // sceAgcAcbAcquireMem
+			"htn36gPnBk4" or // sceAgcAcbWaitRegMem
+			"eZ4+17OQz4Q" or // sceAgcAcbWriteData
+			"j3EtxFkSIhQ" or // sceAgcAcbDispatchIndirect
+			"gSRnr79F8tQ" or // sceAgcDriverSubmitAcb
+			"i1jyy49AjXU" or // sceAgcDcbWriteData
+			"VmW0Tdpy420" or // sceAgcDcbWaitRegMem
+			"WmAc2MEj6Io" or // sceAgcDcbDmaData
+			"rUuVjyR+Rd4" or // sceAgcDcbGetLodStatsGetSize
+			"vuSXe69VILM" or // sceAgcDcbGetLodStats
+			"RmaJwLtc8rY" or // sceAgcDcbSetBaseIndirectArgs
+			"CtB+A9-VxO0" or // sceAgcDcbDispatchIndirect
+			"+kSrjIVxKFE" or // sceAgcDcbPushMarker
+			"H7uZqCoNuWk" or // sceAgcDcbPopMarker
+			"IxYiarKlXxM" or // sceAgcDmaDataPatchSetDstAddressOrOffset
+			"3KDcnM3lrcU" or // sceAgcWaitRegMemPatchAddress
+			"0fWWK5uG9rQ" or // sceAgcQueueEndOfPipeActionPatchAddress
 			"a8uLzYY--tM" or
 			"Qs1xtplKo0U" or
 			"GuchCTefuZw" or
@@ -653,7 +720,18 @@ public sealed partial class DirectExecutionBackend
 			"rqwFKI4PAiM" or
 			"eE4Szl8sil8" or
 			"qvMUCyyaCSI" or
+			"Vo5V8KAwCmk" or // sceSystemServiceHideSplashScreen
+			"TywrFKCoLGY" or // sceSaveDataInitialize3
+			"dyIhnXq-0SM" or // sceSaveDataDirNameSearch
+			"ZP4e7rlzOUk" or // sceSaveDataMount3
+			"ERKzksauAJA" or // sceSaveDataDialogGetStatus
+			"KK3Bdg1RWK0" or // sceSaveDataDialogUpdateStatus
+			"en7gNVnh878" or // sceSaveDataDialogIsReadyToDisplay
+			"jO8DM8oyego" or // sceNpEntitlementAccessInitialize
+			"TFyU+KFBv54" or // sceNpEntitlementAccessGetAddcontEntitlementInfoList
 			"27bAgiJmOh0" or // pthread_cond_timedwait
+			"iQw3iQPhvUQ" or // sceNetCtlCheckCallback
+			"Q2V+iqvjgC0" or // vsnprintf
 			"j4ViWNHEgww" or // strlen
 			"5jNubw4vlAA" or // strnlen
 			"LHMrG7e8G78" or // wcslen

src/SharpEmu.Core/Cpu/Native/DirectExecutionBackend.cs

@@ -347,10 +347,16 @@ public sealed unsafe partial class DirectExecutionBackend : INativeCpuBackend, I
 
 		public string Name { get; init; } = string.Empty;
 
+		public int Priority { get; init; }
+
+		public ulong AffinityMask { get; init; }
+
 		public CpuContext Context { get; init; } = null!;
 
 		public GuestThreadRunState State { get; set; }
 
+		public ulong ExitValue { get; set; }
+
 		public string? BlockReason { get; set; }
 
 		public bool HasBlockedContinuation { get; set; }
@@ -388,14 +394,14 @@ public sealed unsafe partial class DirectExecutionBackend : INativeCpuBackend, I
 		private Action? _work;
 		private volatile bool _stopping;
 
-		public GuestContinuationRunner(ulong guestThreadHandle)
+		public GuestContinuationRunner(ulong guestThreadHandle, ThreadPriority priority)
 		{
 			_guestThreadHandle = guestThreadHandle;
 			_thread = new Thread(ThreadMain)
 			{
 				IsBackground = true,
 				Name = $"GuestContinuation-{guestThreadHandle:X}",
-				Priority = ThreadPriority.BelowNormal,
+				Priority = priority,
 			};
 			_thread.Start();
 		}
@@ -1262,7 +1268,7 @@ public sealed unsafe partial class DirectExecutionBackend : INativeCpuBackend, I
 			   libraryName.IndexOf("Kernel", StringComparison.OrdinalIgnoreCase) >= 0;
 	}
 
-	private static bool PreferLleForLibcExport(string exportName)
+	private bool PreferLleForLibcExport(string exportName)
 	{
 		if (string.IsNullOrWhiteSpace(exportName))
 		{
@@ -1283,6 +1289,10 @@ public sealed unsafe partial class DirectExecutionBackend : INativeCpuBackend, I
 		{
 			return true;
 		}
+		if (IsLibcAllocatorExport(exportName))
+		{
+			return CanUseLleLibcAllocatorFamily();
+		}
 		if (string.Equals(value, "0", StringComparison.Ordinal))
 		{
 			return true;
@@ -1294,6 +1304,51 @@ public sealed unsafe partial class DirectExecutionBackend : INativeCpuBackend, I
 		return IsSafeLleLibcExport(exportName);
 	}
 
+	private bool CanUseLleLibcAllocatorFamily()
+	{
+		return HasUsableLleLibcExport("gQX+4GDQjpM", "malloc") &&
+			   HasUsableLleLibcExport("tIhsqj0qsFE", "free") &&
+			   HasUsableLleLibcExport("2X5agFjKxMc", "calloc") &&
+			   HasUsableLleLibcExport("Y7aJ1uydPMo", "realloc") &&
+			   HasUsableLleLibcExport("Ujf3KzMvRmI", "memalign") &&
+			   HasUsableLleLibcExport("2Btkg8k24Zg", "aligned_alloc") &&
+			   HasUsableLleLibcExport("cVSk9y8URbc", "posix_memalign");
+	}
+
+	private bool HasUsableLleLibcExport(string nid, string exportName)
+	{
+		if (TryResolveRuntimeSymbolAddress(nid, out var address) && IsDirectImportTargetUsable(address))
+		{
+			return true;
+		}
+
+		foreach (var candidate in EnumerateRuntimeSymbolCandidates(exportName))
+		{
+			if (TryResolveRuntimeSymbolAddress(candidate, out address) && IsDirectImportTargetUsable(address))
+			{
+				return true;
+			}
+		}
+
+		return false;
+	}
+
+	private static bool IsLibcAllocatorExport(string exportName)
+	{
+		return exportName switch
+		{
+			"malloc" or
+			"free" or
+			"calloc" or
+			"realloc" or
+			"memalign" or
+			"aligned_alloc" or
+			"posix_memalign" or
+			"malloc_usable_size" => true,
+			_ => false,
+		};
+	}
+
 	private static bool IsSafeLleLibcExport(string exportName)
 	{
 		return exportName switch
@@ -2393,13 +2448,78 @@ public sealed unsafe partial class DirectExecutionBackend : INativeCpuBackend, I
 			Interlocked.Increment(ref _readyGuestThreadCount);
 		}
 		Console.Error.WriteLine(
-			$"[LOADER][INFO] Scheduled guest thread '{thread.Name}' handle=0x{thread.ThreadHandle:X16} entry=0x{thread.EntryPoint:X16} arg=0x{thread.Argument:X16}");
+			$"[LOADER][INFO] Scheduled guest thread '{thread.Name}' handle=0x{thread.ThreadHandle:X16} " +
+			$"entry=0x{thread.EntryPoint:X16} arg=0x{thread.Argument:X16} priority={thread.Priority} " +
+			$"host_priority={MapGuestThreadPriority(thread.Priority)} affinity=0x{thread.AffinityMask:X}");
 		Pump(creatorContext, "pthread_create");
 		return true;
 	}
 
 	public bool SupportsGuestContextTransfer => true;
 
+	public bool TryJoinThread(
+		CpuContext callerContext,
+		ulong threadHandle,
+		out ulong returnValue,
+		out string? error)
+	{
+		returnValue = 0;
+		error = null;
+		if (threadHandle == 0)
+		{
+			error = "thread handle is zero";
+			return false;
+		}
+
+		if (threadHandle == GuestThreadExecution.CurrentGuestThreadHandle)
+		{
+			error = "thread cannot join itself";
+			return false;
+		}
+
+		while (!ActiveForcedGuestExit)
+		{
+			Thread? hostThread;
+			lock (_guestThreadGate)
+			{
+				if (!_guestThreads.TryGetValue(threadHandle, out var thread))
+				{
+					error = $"unknown guest thread 0x{threadHandle:X16}";
+					return false;
+				}
+
+				if (thread.State == GuestThreadRunState.Exited)
+				{
+					returnValue = thread.ExitValue;
+					return true;
+				}
+
+				if (thread.State == GuestThreadRunState.Faulted)
+				{
+					error =
+						$"guest thread 0x{threadHandle:X16} faulted: " +
+						(thread.BlockReason ?? "unknown error");
+					return false;
+				}
+
+				hostThread = thread.HostThread;
+			}
+
+			if (hostThread is not null &&
+				!ReferenceEquals(hostThread, Thread.CurrentThread))
+			{
+				hostThread.Join(1);
+			}
+			else
+			{
+				Thread.Sleep(1);
+			}
+		}
+
+		error = "guest execution stopped while joining thread";
+		return false;
+	}
+
 	public void Pump(CpuContext callerContext, string reason)
 	{
 		_ = callerContext;
@@ -2448,7 +2568,7 @@ public sealed unsafe partial class DirectExecutionBackend : INativeCpuBackend, I
 				{
 					IsBackground = true,
 					Name = $"SharpEmu-{thread.Name}",
-					Priority = ThreadPriority.BelowNormal,
+					Priority = MapGuestThreadPriority(thread.Priority),
 				};
 				lock (_guestThreadGate)
 				{
@@ -2851,7 +2971,9 @@ public sealed unsafe partial class DirectExecutionBackend : INativeCpuBackend, I
 			{
 				if (_guestThreads.TryGetValue(currentGuestThreadHandle, out var guestThread))
 				{
-					runner = guestThread.ContinuationRunner ??= new GuestContinuationRunner(currentGuestThreadHandle);
+					runner = guestThread.ContinuationRunner ??= new GuestContinuationRunner(
+						currentGuestThreadHandle,
+						MapGuestThreadPriority(guestThread.Priority));
 				}
 				else
 				{
@@ -2990,6 +3112,8 @@ public sealed unsafe partial class DirectExecutionBackend : INativeCpuBackend, I
 			EntryPoint = request.EntryPoint,
 			Argument = request.Argument,
 			Name = string.IsNullOrWhiteSpace(request.Name) ? $"Thread-{request.ThreadHandle:X}" : request.Name,
+			Priority = request.Priority,
+			AffinityMask = request.AffinityMask,
 			Context = context,
 			State = GuestThreadRunState.Ready,
 		};
@@ -3129,11 +3253,77 @@ public sealed unsafe partial class DirectExecutionBackend : INativeCpuBackend, I
 			context.TryWriteUInt64(tlsBase + 0x60, tlsBase);
 	}
 
+	private static ThreadPriority MapGuestThreadPriority(int priority)
+	{
+		if (priority <= 478)
+		{
+			return ThreadPriority.Highest;
+		}
+		if (priority >= 733)
+		{
+			return ThreadPriority.Lowest;
+		}
+
+		return ThreadPriority.Normal;
+	}
+
+	private void ApplyGuestThreadAffinity(ulong guestAffinityMask)
+	{
+		var hostAffinityMask = MapGuestThreadAffinity(guestAffinityMask);
+		if (hostAffinityMask == 0)
+		{
+			return;
+		}
+
+		if (SetThreadAffinityMask(GetCurrentThread(), (nuint)hostAffinityMask) == 0 && _logGuestThreads)
+		{
+			Console.Error.WriteLine(
+				$"[LOADER][WARN] Failed to set guest thread affinity guest=0x{guestAffinityMask:X} " +
+				$"host=0x{hostAffinityMask:X} error={Marshal.GetLastWin32Error()}");
+		}
+	}
+
+	private static ulong MapGuestThreadAffinity(ulong guestAffinityMask)
+	{
+		if (guestAffinityMask == 0 || guestAffinityMask == ulong.MaxValue)
+		{
+			return 0;
+		}
+
+		var processorCount = Math.Min(Environment.ProcessorCount, 64);
+		if (processorCount == 0)
+		{
+			return 0;
+		}
+
+		ulong hostAffinityMask = 0;
+		for (var guestCpu = 0; guestCpu < 64; guestCpu++)
+		{
+			if ((guestAffinityMask & (1UL << guestCpu)) == 0)
+			{
+				continue;
+			}
+
+			var hostCpu = processorCount < 8
+				? guestCpu % processorCount
+				: processorCount >= 16
+					? guestCpu * 2
+					: guestCpu;
+			if (hostCpu < processorCount)
+			{
+				hostAffinityMask |= 1UL << hostCpu;
+			}
+		}
+
+		return hostAffinityMask;
+	}
+
 	private void RunGuestThread(GuestThreadState thread, string reason)
 	{
 		var previousLastError = LastError;
 		var previousGuestThreadHandle = GuestThreadExecution.EnterGuestThread(thread.ThreadHandle);
 		var previousGuestThreadState = _activeGuestThreadState;
+		ApplyGuestThreadAffinity(thread.AffinityMask);
 		Volatile.Write(ref thread.HostThreadId, unchecked((int)GetCurrentThreadId()));
 		_activeGuestThreadState = thread;
 		try
@@ -3178,11 +3368,23 @@ public sealed unsafe partial class DirectExecutionBackend : INativeCpuBackend, I
 				switch (exitReason)
 				{
 					case GuestNativeCallExitReason.Returned:
+						thread.ExitValue = thread.Context[CpuRegister.Rax];
 						thread.State = GuestThreadRunState.Exited;
 						break;
 					case GuestNativeCallExitReason.Blocked:
 						thread.State = GuestThreadRunState.Blocked;
 						thread.BlockReason = blockReason;
+						if (thread.HasBlockedContinuation &&
+							thread.BlockWakeHandler is not null &&
+							thread.BlockWakeHandler())
+						{
+							thread.State = GuestThreadRunState.Ready;
+							thread.BlockReason = null;
+							thread.BlockWakeHandler = null;
+							thread.BlockDeadlineTimestamp = 0;
+							_readyGuestThreads.Enqueue(thread);
+							Interlocked.Increment(ref _readyGuestThreadCount);
+						}
 						break;
 					default:
 						thread.State = GuestThreadRunState.Faulted;
@@ -4144,6 +4346,12 @@ public sealed unsafe partial class DirectExecutionBackend : INativeCpuBackend, I
 	[DllImport("kernel32.dll")]
 	private static extern uint GetCurrentThreadId();
 
+	[DllImport("kernel32.dll")]
+	private static extern nint GetCurrentThread();
+
+	[DllImport("kernel32.dll", SetLastError = true)]
+	private static extern nuint SetThreadAffinityMask(nint hThread, nuint dwThreadAffinityMask);
+
 	[DllImport("kernel32.dll", SetLastError = true)]
 	private static extern nint OpenThread(uint dwDesiredAccess, [MarshalAs(UnmanagedType.Bool)] bool bInheritHandle, uint dwThreadId);
 

src/SharpEmu.Core/Memory/PhysicalVirtualMemory.cs

@@ -9,7 +9,7 @@ namespace SharpEmu.Core.Memory;
 
 public sealed unsafe class PhysicalVirtualMemory : IVirtualMemory, IGuestMemoryAllocator, IDisposable
 {
-    private readonly object _gate = new();
+    private readonly ReaderWriterLockSlim _gate = new(LockRecursionPolicy.SupportsRecursion);
     private readonly object _guestAllocationGate = new();
     private readonly List<MemoryRegion> _regions = new();
     private readonly Dictionary<ulong, ProgramHeaderFlags> _pageProtections = new();
@@ -78,7 +78,8 @@ public sealed unsafe class PhysicalVirtualMemory : IVirtualMemory, IGuestMemoryA
             return false;
         }
 
-        lock (_gate)
+        _gate.EnterWriteLock();
+        try
         {
             _regions.Add(new MemoryRegion
             {
@@ -89,6 +90,10 @@ public sealed unsafe class PhysicalVirtualMemory : IVirtualMemory, IGuestMemoryA
                 Protection = protection
             });
         }
+        finally
+        {
+            _gate.ExitWriteLock();
+        }
 
         var allocationKind = executable ? "executable memory" : "data memory";
         Console.Error.WriteLine($"[VMEM] Allocated exact {allocationKind}: 0x{actualAddress:X16} - 0x{actualAddress + alignedSize:X16} ({alignedSize} bytes)");
@@ -202,7 +207,8 @@ public sealed unsafe class PhysicalVirtualMemory : IVirtualMemory, IGuestMemoryA
             }
         }
 
-        lock (_gate)
+        _gate.EnterWriteLock();
+        try
         {
             _regions.Add(new MemoryRegion
             {
@@ -213,6 +219,10 @@ public sealed unsafe class PhysicalVirtualMemory : IVirtualMemory, IGuestMemoryA
                 Protection = protection
             });
         }
+        finally
+        {
+            _gate.ExitWriteLock();
+        }
 
         var allocationKind = reservedOnly
             ? "reserved data memory (lazy commit)"
@@ -315,7 +325,8 @@ public sealed unsafe class PhysicalVirtualMemory : IVirtualMemory, IGuestMemoryA
     {
         lock (_guestAllocationGate)
         {
-            lock (_gate)
+            _gate.EnterWriteLock();
+            try
             {
                 foreach (var region in _regions)
                 {
@@ -324,6 +335,10 @@ public sealed unsafe class PhysicalVirtualMemory : IVirtualMemory, IGuestMemoryA
                 _regions.Clear();
                 _pageProtections.Clear();
             }
+            finally
+            {
+                _gate.ExitWriteLock();
+            }
 
             _guestAllocationArenaBase = 0;
             _guestAllocationOffset = 0;
@@ -343,7 +358,8 @@ public sealed unsafe class PhysicalVirtualMemory : IVirtualMemory, IGuestMemoryA
         var mapEnd = AlignUp(segmentEnd, PageSize);
         var mapSize = checked(mapEnd - mapStart);
 
-        lock (_gate)
+        _gate.EnterWriteLock();
+        try
         {
             var existingRegion = FindRegion(mapStart, mapSize);
             if (existingRegion == null)
@@ -376,6 +392,10 @@ public sealed unsafe class PhysicalVirtualMemory : IVirtualMemory, IGuestMemoryA
 
             Console.Error.WriteLine($"[VMEM] Mapped segment: 0x{virtualAddress:X16} - 0x{virtualAddress + memorySize:X16} (file: {fileData.Length} bytes, prot: {protection})");
         }
+        finally
+        {
+            _gate.ExitWriteLock();
+        }
     }
 
     private void ApplySegmentProtection(ulong mapStart, ulong mapEnd, ProgramHeaderFlags flags)
@@ -425,7 +445,8 @@ public sealed unsafe class PhysicalVirtualMemory : IVirtualMemory, IGuestMemoryA
 
     public IReadOnlyList<VirtualMemoryRegion> SnapshotRegions()
     {
-        lock (_gate)
+        _gate.EnterReadLock();
+        try
         {
             var snapshot = new VirtualMemoryRegion[_regions.Count];
             for (var i = 0; i < _regions.Count; i++)
@@ -440,11 +461,17 @@ public sealed unsafe class PhysicalVirtualMemory : IVirtualMemory, IGuestMemoryA
             }
             return snapshot;
         }
+        finally
+        {
+            _gate.ExitReadLock();
+        }
     }
 
     public bool TryRead(ulong virtualAddress, Span<byte> destination)
     {
-        lock (_gate)
+        var requiresExclusiveAccess = false;
+        _gate.EnterReadLock();
+        try
         {
             foreach (var region in _regions)
             {
@@ -456,48 +483,55 @@ public sealed unsafe class PhysicalVirtualMemory : IVirtualMemory, IGuestMemoryA
                         return true;
                     }
 
-                    if (!EnsureRangeCommitted((ulong)srcPtr, (ulong)destination.Length, region))
+                    if (region.IsReservedOnly)
                     {
-                        return false;
-                    }
-
-                    if (CanReadWithoutProtectionChange((ulong)srcPtr, (ulong)destination.Length, region))
-                    {
-                        fixed (byte* destPtr = destination)
+                        if (!EnsureRangeCommitted((ulong)srcPtr, (ulong)destination.Length, region))
                         {
-                            Buffer.MemoryCopy(srcPtr, destPtr, (nuint)destination.Length, (nuint)destination.Length);
-                        }
-
-                        return true;
-                    }
-
-                    if (!TryTemporarilyProtectForRead((ulong)srcPtr, (ulong)destination.Length, region, out var touchedPages))
-                    {
-                        return false;
-                    }
-
-                    try
-                    {
-                        fixed (byte* destPtr = destination)
-                        {
-                            Buffer.MemoryCopy(srcPtr, destPtr, (nuint)destination.Length, (nuint)destination.Length);
+                            return false;
                         }
                     }
-                    finally
+
+                    if (!CanReadWithoutProtectionChange((ulong)srcPtr, (ulong)destination.Length, region))
                     {
-                        RestorePageProtections(touchedPages);
+                        requiresExclusiveAccess = true;
+                        break;
+                    }
+
+                    fixed (byte* destPtr = destination)
+                    {
+                        Buffer.MemoryCopy(srcPtr, destPtr, (nuint)destination.Length, (nuint)destination.Length);
                     }
 
                     return true;
                 }
             }
+        }
+        finally
+        {
+            _gate.ExitReadLock();
+        }
+
+        if (!requiresExclusiveAccess)
+        {
             return false;
         }
+
+        _gate.EnterWriteLock();
+        try
+        {
+            return TryReadExclusive(virtualAddress, destination);
+        }
+        finally
+        {
+            _gate.ExitWriteLock();
+        }
     }
 
     public bool TryWrite(ulong virtualAddress, ReadOnlySpan<byte> source)
     {
-        lock (_gate)
+        var requiresExclusiveAccess = false;
+        _gate.EnterReadLock();
+        try
         {
             foreach (var region in _regions)
             {
@@ -509,47 +543,148 @@ public sealed unsafe class PhysicalVirtualMemory : IVirtualMemory, IGuestMemoryA
                         return true;
                     }
 
-                    if (!EnsureRangeCommitted((ulong)destPtr, (ulong)source.Length, region))
+                    if (region.IsReservedOnly)
                     {
-                        return false;
-                    }
-
-                    if (CanWriteWithoutProtectionChange((ulong)destPtr, (ulong)source.Length, region))
-                    {
-                        fixed (byte* srcPtr = source)
+                        if (!EnsureRangeCommitted((ulong)destPtr, (ulong)source.Length, region))
                         {
-                            Buffer.MemoryCopy(srcPtr, destPtr, (nuint)source.Length, (nuint)source.Length);
-                        }
-
-                        return true;
-                    }
-
-                    if (!VirtualProtect(destPtr, (nuint)source.Length, PAGE_EXECUTE_READWRITE, out var oldProtect))
-                    {
-                        return false;
-                    }
-
-                    try
-                    {
-                        fixed (byte* srcPtr = source)
-                        {
-                            Buffer.MemoryCopy(srcPtr, destPtr, (nuint)source.Length, (nuint)source.Length);
+                            return false;
                         }
                     }
-                    finally
+
+                    if (!CanWriteWithoutProtectionChange((ulong)destPtr, (ulong)source.Length, region))
                     {
-                        VirtualProtect(destPtr, (nuint)source.Length, oldProtect, out _);
-                        if (IsExecutableProtection(oldProtect))
-                        {
-                            FlushInstructionCache(null, destPtr, (nuint)source.Length);
-                        }
+                        requiresExclusiveAccess = true;
+                        break;
+                    }
+
+                    fixed (byte* srcPtr = source)
+                    {
+                        Buffer.MemoryCopy(srcPtr, destPtr, (nuint)source.Length, (nuint)source.Length);
                     }
 
                     return true;
                 }
             }
+        }
+        finally
+        {
+            _gate.ExitReadLock();
+        }
+
+        if (!requiresExclusiveAccess)
+        {
             return false;
         }
+
+        _gate.EnterWriteLock();
+        try
+        {
+            return TryWriteExclusive(virtualAddress, source);
+        }
+        finally
+        {
+            _gate.ExitWriteLock();
+        }
+    }
+
+    private bool TryReadExclusive(ulong virtualAddress, Span<byte> destination)
+    {
+        foreach (var region in _regions)
+        {
+            if (!TryResolveRegionOffset(virtualAddress, (ulong)destination.Length, region, out var offset))
+            {
+                continue;
+            }
+
+            var srcPtr = (void*)(region.VirtualAddress + offset);
+            if (!EnsureRangeCommitted((ulong)srcPtr, (ulong)destination.Length, region))
+            {
+                return false;
+            }
+
+            if (CanReadWithoutProtectionChange((ulong)srcPtr, (ulong)destination.Length, region))
+            {
+                fixed (byte* destPtr = destination)
+                {
+                    Buffer.MemoryCopy(srcPtr, destPtr, (nuint)destination.Length, (nuint)destination.Length);
+                }
+
+                return true;
+            }
+
+            if (!TryTemporarilyProtectForRead((ulong)srcPtr, (ulong)destination.Length, region, out var touchedPages))
+            {
+                return false;
+            }
+
+            try
+            {
+                fixed (byte* destPtr = destination)
+                {
+                    Buffer.MemoryCopy(srcPtr, destPtr, (nuint)destination.Length, (nuint)destination.Length);
+                }
+            }
+            finally
+            {
+                RestorePageProtections(touchedPages);
+            }
+
+            return true;
+        }
+
+        return false;
+    }
+
+    private bool TryWriteExclusive(ulong virtualAddress, ReadOnlySpan<byte> source)
+    {
+        foreach (var region in _regions)
+        {
+            if (!TryResolveRegionOffset(virtualAddress, (ulong)source.Length, region, out var offset))
+            {
+                continue;
+            }
+
+            var destPtr = (void*)(region.VirtualAddress + offset);
+            if (!EnsureRangeCommitted((ulong)destPtr, (ulong)source.Length, region))
+            {
+                return false;
+            }
+
+            if (CanWriteWithoutProtectionChange((ulong)destPtr, (ulong)source.Length, region))
+            {
+                fixed (byte* srcPtr = source)
+                {
+                    Buffer.MemoryCopy(srcPtr, destPtr, (nuint)source.Length, (nuint)source.Length);
+                }
+
+                return true;
+            }
+
+            if (!VirtualProtect(destPtr, (nuint)source.Length, PAGE_EXECUTE_READWRITE, out var oldProtect))
+            {
+                return false;
+            }
+
+            try
+            {
+                fixed (byte* srcPtr = source)
+                {
+                    Buffer.MemoryCopy(srcPtr, destPtr, (nuint)source.Length, (nuint)source.Length);
+                }
+            }
+            finally
+            {
+                VirtualProtect(destPtr, (nuint)source.Length, oldProtect, out _);
+                if (IsExecutableProtection(oldProtect))
+                {
+                    FlushInstructionCache(null, destPtr, (nuint)source.Length);
+                }
+            }
+
+            return true;
+        }
+
+        return false;
     }
 
     public bool TryWriteUInt64(ulong virtualAddress, ulong value)
@@ -561,7 +696,8 @@ public sealed unsafe class PhysicalVirtualMemory : IVirtualMemory, IGuestMemoryA
 
     public void* GetPointer(ulong virtualAddress)
     {
-        lock (_gate)
+        _gate.EnterReadLock();
+        try
         {
             foreach (var region in _regions)
             {
@@ -573,11 +709,16 @@ public sealed unsafe class PhysicalVirtualMemory : IVirtualMemory, IGuestMemoryA
             }
             return null;
         }
+        finally
+        {
+            _gate.ExitReadLock();
+        }
     }
 
     public bool IsAccessible(ulong virtualAddress, ulong size)
     {
-        lock (_gate)
+        _gate.EnterReadLock();
+        try
         {
             foreach (var region in _regions)
             {
@@ -588,6 +729,10 @@ public sealed unsafe class PhysicalVirtualMemory : IVirtualMemory, IGuestMemoryA
             }
             return false;
         }
+        finally
+        {
+            _gate.ExitReadLock();
+        }
     }
 
     private MemoryRegion? FindRegion(ulong address, ulong size)
@@ -611,7 +756,8 @@ public sealed unsafe class PhysicalVirtualMemory : IVirtualMemory, IGuestMemoryA
         }
 
         var end = address + size;
-        lock (_gate)
+        _gate.EnterReadLock();
+        try
         {
             foreach (var region in _regions)
             {
@@ -622,6 +768,10 @@ public sealed unsafe class PhysicalVirtualMemory : IVirtualMemory, IGuestMemoryA
                 }
             }
         }
+        finally
+        {
+            _gate.ExitReadLock();
+        }
 
         return overlapEnd != 0;
     }
@@ -707,15 +857,26 @@ public sealed unsafe class PhysicalVirtualMemory : IVirtualMemory, IGuestMemoryA
         var endPage = AlignUp(address + size, PageSize);
         var commitProtection = GetCommitProtection(region);
 
-        for (var pageAddress = startPage; pageAddress < endPage; pageAddress += PageSize)
+        var pageAddress = startPage;
+        while (pageAddress < endPage)
         {
             if (VirtualQuery((void*)pageAddress, out var info, (nuint)sizeof(MemoryBasicInformation64)) == 0)
             {
                 return false;
             }
 
+            var queriedEnd = info.RegionSize > ulong.MaxValue - info.BaseAddress
+                ? ulong.MaxValue
+                : info.BaseAddress + info.RegionSize;
+            var rangeEnd = Math.Min(endPage, queriedEnd);
+            if (rangeEnd <= pageAddress)
+            {
+                return false;
+            }
+
             if (info.State == MEM_COMMIT)
             {
+                pageAddress = rangeEnd;
                 continue;
             }
 
@@ -724,10 +885,13 @@ public sealed unsafe class PhysicalVirtualMemory : IVirtualMemory, IGuestMemoryA
                 return false;
             }
 
-            if (VirtualAlloc((void*)pageAddress, (nuint)PageSize, MEM_COMMIT, commitProtection) == null)
+            var commitSize = rangeEnd - pageAddress;
+            if (VirtualAlloc((void*)pageAddress, (nuint)commitSize, MEM_COMMIT, commitProtection) == null)
             {
                 return false;
             }
+
+            pageAddress = rangeEnd;
         }
 
         return true;

src/SharpEmu.Core/Runtime/SharpEmuRuntime.cs

@@ -10,6 +10,7 @@ using SharpEmu.HLE;
 using SharpEmu.Libs.VideoOut;
 using SharpEmu.Libs.Kernel;
 using SharpEmu.Libs.AppContent;
+using SharpEmu.Libs.SaveData;
 using System.Buffers.Binary;
 using System.Collections.Generic;
 using System.Linq;
@@ -136,6 +137,7 @@ public sealed class SharpEmuRuntime : ISharpEmuRuntime
         KernelModuleRegistry.Reset();
         var image = LoadImage(normalizedEbootPath);
         VideoOutExports.ConfigureApplicationInfo(image.Title, image.TitleId, image.Version);
+        SaveDataExports.ConfigureApplicationInfo(image.TitleId);
         RegisterLoadedModule(normalizedEbootPath, image, isMain: true, isSystemModule: false);
         KernelRuntimeCompatExports.ConfigureProcessProcParamAddress(image.ProcParamAddress);
         Console.Error.WriteLine($"[RUNTIME] Entry: 0x{image.EntryPoint:X16}");

src/SharpEmu.HLE/GuestThreadExecution.cs

@@ -10,7 +10,9 @@ public readonly record struct GuestThreadStartRequest(
     ulong EntryPoint,
     ulong Argument,
     ulong AttributeAddress,
-    string Name);
+    string Name,
+    int Priority,
+    ulong AffinityMask);
 
 public readonly record struct GuestThreadSnapshot(
     ulong ThreadHandle,
@@ -27,6 +29,12 @@ public interface IGuestThreadScheduler
 
     bool TryStartThread(CpuContext creatorContext, GuestThreadStartRequest request, out string? error);
 
+    bool TryJoinThread(
+        CpuContext callerContext,
+        ulong threadHandle,
+        out ulong returnValue,
+        out string? error);
+
     void Pump(CpuContext callerContext, string reason);
 
     int WakeBlockedThreads(string wakeKey, int maxCount = int.MaxValue);

src/SharpEmu.Libs/Agc/Gen5ShaderIr.cs

@@ -0,0 +1,313 @@
+// Copyright (C) 2026 SharpEmu Emulator Project
+// SPDX-License-Identifier: GPL-2.0-or-later
+
+namespace SharpEmu.Libs.Agc;
+
+internal enum Gen5ShaderEncoding
+{
+    Sop1,
+    Sop2,
+    Sopc,
+    Sopp,
+    Sopk,
+    Smrd,
+    Smem,
+    Mubuf,
+    Mtbuf,
+    Vop1,
+    Vop2,
+    Vopc,
+    Vop3,
+    Vintrp,
+    Ds,
+    Flat,
+    Vop3p,
+    Mimg,
+    Exp,
+}
+
+internal enum Gen5OperandKind
+{
+    ScalarRegister,
+    VectorRegister,
+    EncodedConstant,
+    LiteralConstant,
+}
+
+internal enum Gen5ShaderResourceKind
+{
+    ReadOnlyTexture,
+    ReadWriteTexture,
+    Sampler,
+    ConstantBuffer,
+}
+
+internal enum Gen5PixelOutputKind
+{
+    Float,
+    Uint,
+    Sint,
+}
+
+internal enum Gen5SpirvStage
+{
+    Vertex,
+    Pixel,
+    Compute,
+}
+
+internal sealed record Gen5SpirvShader(
+    byte[] Spirv,
+    IReadOnlyList<Gen5GlobalMemoryBinding> GlobalMemoryBindings,
+    IReadOnlyList<Gen5ImageBinding> ImageBindings,
+    uint AttributeCount,
+    IReadOnlyList<Gen5VertexInputBinding> VertexInputs);
+
+internal readonly record struct Gen5ShaderResourceMapping(
+    Gen5ShaderResourceKind Kind,
+    uint Slot,
+    uint OffsetDwords,
+    bool SizeFlag);
+
+internal sealed record Gen5ShaderMetadata(
+    uint ExtendedUserDataSizeDwords,
+    uint ShaderResourceTableSizeDwords,
+    IReadOnlyDictionary<uint, uint> DirectResources,
+    IReadOnlyList<Gen5ShaderResourceMapping> Resources);
+
+internal readonly record struct Gen5ComputeSystemRegisters(
+    uint? WorkGroupXRegister,
+    uint? WorkGroupYRegister,
+    uint? WorkGroupZRegister,
+    uint? ThreadGroupSizeRegister)
+{
+    public bool TryGetExpression(uint scalarRegister, out string expression)
+    {
+        if (WorkGroupXRegister == scalarRegister)
+        {
+            expression = "gl_WorkGroupID.x";
+            return true;
+        }
+
+        if (WorkGroupYRegister == scalarRegister)
+        {
+            expression = "gl_WorkGroupID.y";
+            return true;
+        }
+
+        if (WorkGroupZRegister == scalarRegister)
+        {
+            expression = "gl_WorkGroupID.z";
+            return true;
+        }
+
+        if (ThreadGroupSizeRegister == scalarRegister)
+        {
+            expression = "(gl_WorkGroupSize.x * gl_WorkGroupSize.y * gl_WorkGroupSize.z)";
+            return true;
+        }
+
+        expression = string.Empty;
+        return false;
+    }
+
+    public void ClearStaticValues(Span<uint> scalarRegisters)
+    {
+        ClearStaticValue(scalarRegisters, WorkGroupXRegister);
+        ClearStaticValue(scalarRegisters, WorkGroupYRegister);
+        ClearStaticValue(scalarRegisters, WorkGroupZRegister);
+        ClearStaticValue(scalarRegisters, ThreadGroupSizeRegister);
+    }
+
+    private static void ClearStaticValue(Span<uint> scalarRegisters, uint? scalarRegister)
+    {
+        if (scalarRegister is { } register && register < scalarRegisters.Length)
+        {
+            scalarRegisters[(int)register] = 0;
+        }
+    }
+}
+
+internal sealed record Gen5ShaderState(
+    Gen5ShaderProgram Program,
+    IReadOnlyList<uint> UserData,
+    Gen5ShaderMetadata? Metadata,
+    Gen5ComputeSystemRegisters? ComputeSystemRegisters = null,
+    uint UserDataScalarRegisterBase = 0);
+
+internal readonly record struct Gen5Operand(Gen5OperandKind Kind, uint Value)
+{
+    public static Gen5Operand Scalar(uint index) =>
+        new(Gen5OperandKind.ScalarRegister, index);
+
+    public static Gen5Operand Vector(uint index) =>
+        new(Gen5OperandKind.VectorRegister, index);
+
+    public static Gen5Operand Source(uint encoded, uint? literal = null)
+    {
+        if (encoded >= 256)
+        {
+            return Vector(encoded - 256);
+        }
+
+        if (encoded is 249 or 255 && literal.HasValue)
+        {
+            return new(Gen5OperandKind.LiteralConstant, literal.Value);
+        }
+
+        if (encoded <= 105 || encoded is 106 or 107 or 124 or 126 or 127)
+        {
+            return Scalar(encoded);
+        }
+
+        return new(Gen5OperandKind.EncodedConstant, encoded);
+    }
+
+    public override string ToString() => Kind switch
+    {
+        Gen5OperandKind.ScalarRegister => $"s{Value}",
+        Gen5OperandKind.VectorRegister => $"v{Value}",
+        Gen5OperandKind.LiteralConstant => $"0x{Value:X8}",
+        _ => $"src[{Value}]",
+    };
+}
+
+internal abstract record Gen5InstructionControl;
+
+internal sealed record Gen5ImageControl(
+    uint Dmask,
+    uint VectorAddress,
+    IReadOnlyList<uint> AddressRegisters,
+    uint VectorData,
+    uint ScalarResource,
+    uint ScalarSampler,
+    uint Dimension,
+    bool IsArray,
+    bool Glc,
+    bool Slc) : Gen5InstructionControl
+{
+    public uint GetAddressRegister(int component) =>
+        component < AddressRegisters.Count
+            ? AddressRegisters[component]
+            : VectorAddress + (uint)component;
+}
+
+internal sealed record Gen5GlobalMemoryControl(
+    uint DwordCount,
+    uint VectorAddress,
+    uint VectorData,
+    uint ScalarAddress,
+    int OffsetBytes,
+    bool Glc,
+    bool Slc) : Gen5InstructionControl;
+
+internal sealed record Gen5BufferMemoryControl(
+    uint DwordCount,
+    uint VectorAddress,
+    uint VectorData,
+    uint ScalarResource,
+    int OffsetBytes,
+    bool IndexEnabled,
+    bool OffsetEnabled,
+    bool Glc,
+    bool Slc) : Gen5InstructionControl;
+
+internal sealed record Gen5ExportControl(
+    uint Target,
+    uint EnableMask,
+    bool Compressed,
+    bool Done,
+    bool ValidMask) : Gen5InstructionControl;
+
+internal sealed record Gen5InterpolationControl(
+    uint Attribute,
+    uint Channel) : Gen5InstructionControl;
+
+internal sealed record Gen5Vop3Control(
+    uint AbsoluteMask,
+    uint NegateMask,
+    uint OutputModifier,
+    bool Clamp,
+    uint? ScalarDestination) : Gen5InstructionControl;
+
+internal sealed record Gen5SdwaControl(
+    uint DestinationSelect,
+    uint Source0Select,
+    uint Source1Select,
+    uint AbsoluteMask,
+    uint NegateMask,
+    uint OutputModifier,
+    bool Clamp) : Gen5InstructionControl;
+
+internal sealed record Gen5DppControl(
+    uint Control,
+    bool FetchInactive,
+    bool BoundControl,
+    uint AbsoluteMask,
+    uint NegateMask,
+    uint BankMask,
+    uint RowMask) : Gen5InstructionControl;
+
+internal sealed record Gen5ScalarMemoryControl(
+    uint DestinationCount,
+    int ImmediateOffsetBytes,
+    uint? DynamicOffsetRegister) : Gen5InstructionControl;
+
+internal sealed record Gen5DataShareControl(
+    uint Offset0,
+    uint Offset1,
+    bool Gds) : Gen5InstructionControl;
+
+internal sealed record Gen5ImageBinding(
+    uint Pc,
+    string Opcode,
+    Gen5ImageControl Control,
+    IReadOnlyList<uint> ResourceDescriptor,
+    IReadOnlyList<uint> SamplerDescriptor,
+    uint? MipLevel);
+
+internal sealed record Gen5GlobalMemoryBinding(
+    uint ScalarAddress,
+    ulong BaseAddress,
+    IReadOnlyList<uint> InstructionPcs,
+    byte[] Data);
+
+internal sealed record Gen5VertexInputBinding(
+    uint Pc,
+    uint Location,
+    uint ComponentCount,
+    uint DataFormat,
+    uint NumberFormat,
+    ulong BaseAddress,
+    uint Stride,
+    uint OffsetBytes,
+    byte[] Data);
+
+internal sealed record Gen5ShaderEvaluation(
+    IReadOnlyList<uint> InitialScalarRegisters,
+    IReadOnlyList<uint> ScalarRegisters,
+    IReadOnlyDictionary<uint, IReadOnlyList<uint>> ScalarRegistersByPc,
+    IReadOnlyList<Gen5ImageBinding> ImageBindings,
+    IReadOnlyList<Gen5GlobalMemoryBinding> GlobalMemoryBindings,
+    Gen5ComputeSystemRegisters? ComputeSystemRegisters = null,
+    IReadOnlySet<uint>? RuntimeScalarRegisters = null,
+    IReadOnlyList<Gen5VertexInputBinding>? VertexInputs = null);
+
+internal sealed record Gen5ShaderInstruction(
+    uint Pc,
+    Gen5ShaderEncoding Encoding,
+    string Opcode,
+    IReadOnlyList<uint> Words,
+    IReadOnlyList<Gen5Operand> Sources,
+    IReadOnlyList<Gen5Operand> Destinations,
+    Gen5InstructionControl? Control);
+
+internal sealed record Gen5ShaderProgram(
+    ulong Address,
+    IReadOnlyList<Gen5ShaderInstruction> Instructions)
+{
+    public IEnumerable<Gen5ImageControl> ImageResources =>
+        Instructions
+            .Select(instruction => instruction.Control)
+            .OfType<Gen5ImageControl>();
+}

src/SharpEmu.Libs/Agc/Gen5ShaderMetadataReader.cs

@@ -0,0 +1,154 @@
+// Copyright (C) 2026 SharpEmu Emulator Project
+// SPDX-License-Identifier: GPL-2.0-or-later
+
+using SharpEmu.HLE;
+using System.Buffers.Binary;
+
+namespace SharpEmu.Libs.Agc;
+
+internal static class Gen5ShaderMetadataReader
+{
+    private const ulong ShaderUserDataOffset = 0x08;
+    private const int ResourceClassCount = 4;
+    private const int MaxMetadataEntries = 4096;
+
+    public static bool TryRead(
+        CpuContext ctx,
+        ulong shaderHeaderAddress,
+        out Gen5ShaderMetadata metadata)
+    {
+        metadata = default!;
+        if (!TryReadUInt64(ctx, shaderHeaderAddress + ShaderUserDataOffset, out var userDataAddress) ||
+            userDataAddress == 0 ||
+            !TryReadUInt64(ctx, userDataAddress, out var directResourceOffsetsAddress))
+        {
+            return false;
+        }
+
+        var resourceOffsets = new ulong[ResourceClassCount];
+        for (var resourceClass = 0; resourceClass < ResourceClassCount; resourceClass++)
+        {
+            if (!TryReadUInt64(
+                    ctx,
+                    userDataAddress + 0x08 + (ulong)(resourceClass * sizeof(ulong)),
+                    out resourceOffsets[resourceClass]))
+            {
+                return false;
+            }
+        }
+
+        if (!TryReadUInt16(ctx, userDataAddress + 0x28, out var extendedUserDataSize) ||
+            !TryReadUInt16(ctx, userDataAddress + 0x2A, out var shaderResourceTableSize) ||
+            !TryReadUInt16(ctx, userDataAddress + 0x2C, out var directResourceCount) ||
+            directResourceCount > MaxMetadataEntries)
+        {
+            return false;
+        }
+
+        var resourceCounts = new ushort[ResourceClassCount];
+        for (var resourceClass = 0; resourceClass < ResourceClassCount; resourceClass++)
+        {
+            if (!TryReadUInt16(
+                    ctx,
+                    userDataAddress + 0x2E + (ulong)(resourceClass * sizeof(ushort)),
+                    out resourceCounts[resourceClass]) ||
+                resourceCounts[resourceClass] > MaxMetadataEntries)
+            {
+                return false;
+            }
+        }
+
+        var directResources = new Dictionary<uint, uint>();
+        if (directResourceCount != 0)
+        {
+            if (directResourceOffsetsAddress == 0)
+            {
+                return false;
+            }
+
+            for (uint type = 0; type < directResourceCount; type++)
+            {
+                if (!TryReadUInt16(ctx, directResourceOffsetsAddress + type * sizeof(ushort), out var offset))
+                {
+                    return false;
+                }
+
+                if (offset != ushort.MaxValue)
+                {
+                    directResources[type] = offset;
+                }
+            }
+        }
+
+        var resources = new List<Gen5ShaderResourceMapping>();
+        for (var resourceClass = 0; resourceClass < ResourceClassCount; resourceClass++)
+        {
+            var count = resourceCounts[resourceClass];
+            if (count == 0)
+            {
+                continue;
+            }
+
+            if (resourceOffsets[resourceClass] == 0)
+            {
+                return false;
+            }
+
+            for (uint slot = 0; slot < count; slot++)
+            {
+                if (!TryReadUInt16(
+                        ctx,
+                        resourceOffsets[resourceClass] + slot * sizeof(ushort),
+                        out var sharp))
+                {
+                    return false;
+                }
+
+                var offset = (uint)(sharp & 0x7FFF);
+                if (offset == 0x7FFF)
+                {
+                    continue;
+                }
+
+                resources.Add(new Gen5ShaderResourceMapping(
+                    (Gen5ShaderResourceKind)resourceClass,
+                    slot,
+                    offset,
+                    (sharp & 0x8000) != 0));
+            }
+        }
+
+        metadata = new Gen5ShaderMetadata(
+            extendedUserDataSize,
+            shaderResourceTableSize,
+            directResources,
+            resources);
+        return true;
+    }
+
+    private static bool TryReadUInt16(CpuContext ctx, ulong address, out ushort value)
+    {
+        Span<byte> bytes = stackalloc byte[sizeof(ushort)];
+        if (!ctx.Memory.TryRead(address, bytes))
+        {
+            value = 0;
+            return false;
+        }
+
+        value = BinaryPrimitives.ReadUInt16LittleEndian(bytes);
+        return true;
+    }
+
+    private static bool TryReadUInt64(CpuContext ctx, ulong address, out ulong value)
+    {
+        Span<byte> bytes = stackalloc byte[sizeof(ulong)];
+        if (!ctx.Memory.TryRead(address, bytes))
+        {
+            value = 0;
+            return false;
+        }
+
+        value = BinaryPrimitives.ReadUInt64LittleEndian(bytes);
+        return true;
+    }
+}

src/SharpEmu.Libs/Agc/SpirvFixedShaders.cs

@@ -0,0 +1,167 @@
+// Copyright (C) 2026 SharpEmu Emulator Project
+// SPDX-License-Identifier: GPL-2.0-or-later
+
+namespace SharpEmu.Libs.Agc;
+
+internal static class SpirvFixedShaders
+{
+    public static byte[] CreateFullscreenVertex(uint attributeCount)
+    {
+        var module = new SpirvModuleBuilder();
+        module.AddCapability(SpirvCapability.Shader);
+
+        var voidType = module.TypeVoid();
+        var boolType = module.TypeBool();
+        var uintType = module.TypeInt(32, signed: false);
+        var floatType = module.TypeFloat(32);
+        var vec4Type = module.TypeVector(floatType, 4);
+        var inputUintPointer = module.TypePointer(SpirvStorageClass.Input, uintType);
+        var outputVec4Pointer = module.TypePointer(SpirvStorageClass.Output, vec4Type);
+
+        var vertexIndex = module.AddGlobalVariable(inputUintPointer, SpirvStorageClass.Input);
+        module.AddName(vertexIndex, "vertexIndex");
+        module.AddDecoration(
+            vertexIndex,
+            SpirvDecoration.BuiltIn,
+            (uint)SpirvBuiltIn.VertexIndex);
+
+        var position = module.AddGlobalVariable(outputVec4Pointer, SpirvStorageClass.Output);
+        module.AddName(position, "position");
+        module.AddDecoration(position, SpirvDecoration.BuiltIn, (uint)SpirvBuiltIn.Position);
+
+        var attributes = new uint[attributeCount];
+        for (uint index = 0; index < attributeCount; index++)
+        {
+            attributes[index] =
+                module.AddGlobalVariable(outputVec4Pointer, SpirvStorageClass.Output);
+            module.AddName(attributes[index], $"attr{index}");
+            module.AddDecoration(attributes[index], SpirvDecoration.Location, index);
+            module.AddDecoration(attributes[index], SpirvDecoration.NoPerspective);
+        }
+
+        var functionType = module.TypeFunction(voidType);
+        var main = module.BeginFunction(voidType, functionType);
+        module.AddName(main, "main");
+        module.AddLabel();
+
+        var indexValue = module.AddInstruction(SpirvOp.Load, uintType, vertexIndex);
+        var one = module.Constant(uintType, 1);
+        var two = module.Constant(uintType, 2);
+        var shifted = module.AddInstruction(SpirvOp.ShiftLeftLogical, uintType, indexValue, one);
+        var xBits = module.AddInstruction(SpirvOp.BitwiseAnd, uintType, shifted, two);
+        var yBits = module.AddInstruction(SpirvOp.BitwiseAnd, uintType, indexValue, two);
+        var x = module.AddInstruction(SpirvOp.ConvertUToF, floatType, xBits);
+        var y = module.AddInstruction(SpirvOp.ConvertUToF, floatType, yBits);
+        var zero = module.ConstantFloat(floatType, 0f);
+        var oneFloat = module.ConstantFloat(floatType, 1f);
+        var twoFloat = module.ConstantFloat(floatType, 2f);
+        var xPosition = module.AddInstruction(SpirvOp.FMul, floatType, x, twoFloat);
+        xPosition = module.AddInstruction(SpirvOp.FSub, floatType, xPosition, oneFloat);
+        var yPosition = module.AddInstruction(SpirvOp.FMul, floatType, y, twoFloat);
+        yPosition = module.AddInstruction(SpirvOp.FSub, floatType, yPosition, oneFloat);
+        var positionValue = module.AddInstruction(
+            SpirvOp.CompositeConstruct,
+            vec4Type,
+            xPosition,
+            yPosition,
+            zero,
+            oneFloat);
+        module.AddStatement(SpirvOp.Store, position, positionValue);
+
+        var attributeValue = module.AddInstruction(
+            SpirvOp.CompositeConstruct,
+            vec4Type,
+            x,
+            y,
+            zero,
+            oneFloat);
+        foreach (var attribute in attributes)
+        {
+            module.AddStatement(SpirvOp.Store, attribute, attributeValue);
+        }
+
+        module.AddStatement(SpirvOp.Return);
+        module.EndFunction();
+
+        var interfaces = new uint[2 + attributes.Length];
+        interfaces[0] = vertexIndex;
+        interfaces[1] = position;
+        attributes.CopyTo(interfaces, 2);
+        module.AddEntryPoint(SpirvExecutionModel.Vertex, main, "main", interfaces);
+        _ = boolType;
+        return module.Build();
+    }
+
+    public static byte[] CreateCopyFragment()
+    {
+        var module = new SpirvModuleBuilder();
+        module.AddCapability(SpirvCapability.Shader);
+
+        var voidType = module.TypeVoid();
+        var floatType = module.TypeFloat(32);
+        var vec2Type = module.TypeVector(floatType, 2);
+        var vec4Type = module.TypeVector(floatType, 4);
+        var inputVec4Pointer = module.TypePointer(SpirvStorageClass.Input, vec4Type);
+        var outputVec4Pointer = module.TypePointer(SpirvStorageClass.Output, vec4Type);
+        var imageType = module.TypeImage(
+            floatType,
+            SpirvImageDim.Dim2D,
+            depth: false,
+            arrayed: false,
+            multisampled: false,
+            sampled: 1,
+            SpirvImageFormat.Unknown);
+        var sampledImageType = module.TypeSampledImage(imageType);
+        var sampledImagePointer =
+            module.TypePointer(SpirvStorageClass.UniformConstant, sampledImageType);
+
+        var attribute = module.AddGlobalVariable(inputVec4Pointer, SpirvStorageClass.Input);
+        module.AddName(attribute, "attr0");
+        module.AddDecoration(attribute, SpirvDecoration.Location, 0);
+
+        var texture = module.AddGlobalVariable(
+            sampledImagePointer,
+            SpirvStorageClass.UniformConstant);
+        module.AddName(texture, "tex0");
+        module.AddDecoration(texture, SpirvDecoration.DescriptorSet, 0);
+        module.AddDecoration(texture, SpirvDecoration.Binding, 1);
+
+        var output = module.AddGlobalVariable(outputVec4Pointer, SpirvStorageClass.Output);
+        module.AddName(output, "outColor");
+        module.AddDecoration(output, SpirvDecoration.Location, 0);
+
+        var functionType = module.TypeFunction(voidType);
+        var main = module.BeginFunction(voidType, functionType);
+        module.AddName(main, "main");
+        module.AddLabel();
+
+        var attributeValue = module.AddInstruction(SpirvOp.Load, vec4Type, attribute);
+        var coordinates = module.AddInstruction(
+            SpirvOp.VectorShuffle,
+            vec2Type,
+            attributeValue,
+            attributeValue,
+            0,
+            1);
+        var sampledImage = module.AddInstruction(SpirvOp.Load, sampledImageType, texture);
+        var lod = module.ConstantFloat(floatType, 0f);
+        var color = module.AddInstruction(
+            SpirvOp.ImageSampleExplicitLod,
+            vec4Type,
+            sampledImage,
+            coordinates,
+            2,
+            lod);
+        module.AddStatement(SpirvOp.Store, output, color);
+        module.AddStatement(SpirvOp.Return);
+        module.EndFunction();
+
+        module.AddEntryPoint(
+            SpirvExecutionModel.Fragment,
+            main,
+            "main",
+            [attribute, texture, output]);
+        module.AddExecutionMode(main, SpirvExecutionMode.OriginUpperLeft);
+        return module.Build();
+    }
+}

src/SharpEmu.Libs/Agc/SpirvModuleBuilder.cs

@@ -0,0 +1,885 @@
+// Copyright (C) 2026 SharpEmu Emulator Project
+// SPDX-License-Identifier: GPL-2.0-or-later
+
+using System.Buffers.Binary;
+using System.Text;
+
+namespace SharpEmu.Libs.Agc;
+
+internal enum SpirvOp : ushort
+{
+    Nop = 0,
+    Name = 5,
+    Extension = 10,
+    ExtInstImport = 11,
+    ExtInst = 12,
+    MemoryModel = 14,
+    EntryPoint = 15,
+    ExecutionMode = 16,
+    Capability = 17,
+    TypeVoid = 19,
+    TypeBool = 20,
+    TypeInt = 21,
+    TypeFloat = 22,
+    TypeVector = 23,
+    TypeImage = 25,
+    TypeSampler = 26,
+    TypeSampledImage = 27,
+    TypeArray = 28,
+    TypeRuntimeArray = 29,
+    TypeStruct = 30,
+    TypePointer = 32,
+    TypeFunction = 33,
+    ConstantTrue = 41,
+    ConstantFalse = 42,
+    Constant = 43,
+    ConstantComposite = 44,
+    ConstantNull = 46,
+    Function = 54,
+    FunctionParameter = 55,
+    FunctionEnd = 56,
+    FunctionCall = 57,
+    Variable = 59,
+    Load = 61,
+    Store = 62,
+    AccessChain = 65,
+    ArrayLength = 68,
+    Decorate = 71,
+    VectorExtractDynamic = 77,
+    VectorInsertDynamic = 78,
+    VectorShuffle = 79,
+    CompositeConstruct = 80,
+    CompositeExtract = 81,
+    CompositeInsert = 82,
+    CopyObject = 83,
+    SampledImage = 86,
+    ImageSampleImplicitLod = 87,
+    ImageSampleExplicitLod = 88,
+    ImageSampleDrefImplicitLod = 89,
+    ImageSampleDrefExplicitLod = 90,
+    ImageFetch = 95,
+    ImageGather = 96,
+    ImageDrefGather = 97,
+    ImageRead = 98,
+    ImageWrite = 99,
+    Image = 100,
+    ImageQuerySizeLod = 103,
+    ImageQuerySize = 104,
+    ImageQueryLod = 105,
+    ImageQueryLevels = 106,
+    ImageQuerySamples = 107,
+    ConvertFToU = 109,
+    ConvertFToS = 110,
+    ConvertSToF = 111,
+    ConvertUToF = 112,
+    UConvert = 113,
+    SConvert = 114,
+    FConvert = 115,
+    Bitcast = 124,
+    SNegate = 126,
+    FNegate = 127,
+    IAdd = 128,
+    FAdd = 129,
+    ISub = 130,
+    FSub = 131,
+    IMul = 132,
+    FMul = 133,
+    UDiv = 134,
+    SDiv = 135,
+    FDiv = 136,
+    UMod = 137,
+    SRem = 138,
+    SMod = 139,
+    FRem = 140,
+    FMod = 141,
+    IAddCarry = 149,
+    ISubBorrow = 150,
+    UMulExtended = 151,
+    SMulExtended = 152,
+    Any = 154,
+    All = 155,
+    IsNan = 156,
+    IsInf = 157,
+    LogicalEqual = 164,
+    LogicalNotEqual = 165,
+    LogicalOr = 166,
+    LogicalAnd = 167,
+    LogicalNot = 168,
+    Select = 169,
+    IEqual = 170,
+    INotEqual = 171,
+    UGreaterThan = 172,
+    SGreaterThan = 173,
+    UGreaterThanEqual = 174,
+    SGreaterThanEqual = 175,
+    ULessThan = 176,
+    SLessThan = 177,
+    ULessThanEqual = 178,
+    SLessThanEqual = 179,
+    FOrdEqual = 180,
+    FUnordEqual = 181,
+    FOrdNotEqual = 182,
+    FUnordNotEqual = 183,
+    FOrdLessThan = 184,
+    FUnordLessThan = 185,
+    FOrdGreaterThan = 186,
+    FUnordGreaterThan = 187,
+    FOrdLessThanEqual = 188,
+    FUnordLessThanEqual = 189,
+    FOrdGreaterThanEqual = 190,
+    FUnordGreaterThanEqual = 191,
+    ShiftRightLogical = 194,
+    ShiftRightArithmetic = 195,
+    ShiftLeftLogical = 196,
+    BitwiseOr = 197,
+    BitwiseXor = 198,
+    BitwiseAnd = 199,
+    Not = 200,
+    BitFieldInsert = 201,
+    BitFieldSExtract = 202,
+    BitFieldUExtract = 203,
+    BitReverse = 204,
+    BitCount = 205,
+    ControlBarrier = 224,
+    MemoryBarrier = 225,
+    AtomicIAdd = 234,
+    Phi = 245,
+    LoopMerge = 246,
+    SelectionMerge = 247,
+    Label = 248,
+    Branch = 249,
+    BranchConditional = 250,
+    Switch = 251,
+    Kill = 252,
+    Return = 253,
+    ReturnValue = 254,
+    Unreachable = 255,
+    GroupNonUniformElect = 333,
+    GroupNonUniformAll = 334,
+    GroupNonUniformAny = 335,
+    GroupNonUniformAllEqual = 336,
+    GroupNonUniformBroadcast = 337,
+    GroupNonUniformBroadcastFirst = 338,
+    GroupNonUniformBallot = 339,
+    GroupNonUniformShuffle = 345,
+    GroupNonUniformShuffleXor = 346,
+    GroupNonUniformShuffleUp = 347,
+    GroupNonUniformShuffleDown = 348,
+}
+
+internal enum SpirvCapability : uint
+{
+    Shader = 1,
+    Float16 = 9,
+    Float64 = 10,
+    Int64 = 11,
+    Int16 = 22,
+    ImageGatherExtended = 25,
+    StorageImageExtendedFormats = 49,
+    ImageQuery = 50,
+    StorageImageReadWithoutFormat = 55,
+    StorageImageWriteWithoutFormat = 56,
+    GroupNonUniform = 61,
+    GroupNonUniformVote = 62,
+    GroupNonUniformBallot = 64,
+    GroupNonUniformShuffle = 65,
+    RuntimeDescriptorArray = 5302,
+}
+
+internal enum SpirvStorageClass : uint
+{
+    UniformConstant = 0,
+    Input = 1,
+    Uniform = 2,
+    Output = 3,
+    Workgroup = 4,
+    Private = 6,
+    Function = 7,
+    PushConstant = 9,
+    Image = 11,
+    StorageBuffer = 12,
+}
+
+internal enum SpirvExecutionModel : uint
+{
+    Vertex = 0,
+    Fragment = 4,
+    GLCompute = 5,
+}
+
+internal enum SpirvExecutionMode : uint
+{
+    OriginUpperLeft = 7,
+    DepthReplacing = 12,
+    LocalSize = 17,
+}
+
+internal enum SpirvDecoration : uint
+{
+    Block = 2,
+    ArrayStride = 6,
+    BuiltIn = 11,
+    NoPerspective = 13,
+    Flat = 14,
+    Location = 30,
+    Binding = 33,
+    DescriptorSet = 34,
+    Offset = 35,
+}
+
+internal enum SpirvBuiltIn : uint
+{
+    Position = 0,
+    VertexIndex = 42,
+    InstanceIndex = 43,
+    FragCoord = 15,
+    FrontFacing = 17,
+    WorkgroupId = 26,
+    LocalInvocationId = 27,
+    GlobalInvocationId = 28,
+    LocalInvocationIndex = 29,
+    SubgroupLocalInvocationId = 41,
+}
+
+internal enum SpirvImageDim : uint
+{
+    Dim1D = 0,
+    Dim2D = 1,
+    Dim3D = 2,
+    Cube = 3,
+    Buffer = 5,
+}
+
+internal enum SpirvImageFormat : uint
+{
+    Unknown = 0,
+    Rgba32f = 1,
+    Rgba16f = 2,
+    R32f = 3,
+    Rgba8 = 4,
+    Rgba8Snorm = 5,
+    Rg32f = 6,
+    Rg16f = 7,
+    R11fG11fB10f = 8,
+    R16f = 9,
+    Rgba16 = 10,
+    Rgb10A2 = 11,
+    Rg16 = 12,
+    Rg8 = 13,
+    R16 = 14,
+    R8 = 15,
+    Rgba16Snorm = 16,
+    Rg16Snorm = 17,
+    Rg8Snorm = 18,
+    R16Snorm = 19,
+    R8Snorm = 20,
+    Rgba32i = 21,
+    Rgba16i = 22,
+    Rgba8i = 23,
+    R32i = 24,
+    Rg32i = 25,
+    Rg16i = 26,
+    Rg8i = 27,
+    R16i = 28,
+    R8i = 29,
+    Rgba32ui = 30,
+    Rgba16ui = 31,
+    Rgba8ui = 32,
+    R32ui = 33,
+    Rgb10A2ui = 34,
+    Rg32ui = 35,
+    Rg16ui = 36,
+    Rg8ui = 37,
+    R16ui = 38,
+    R8ui = 39,
+}
+
+internal sealed class SpirvModuleBuilder
+{
+    private const uint Magic = 0x07230203;
+    private const uint Version15 = 0x00010500;
+    private const uint Generator = 0x53504500; // "SPE"
+
+    private readonly List<uint> _capabilities = [];
+    private readonly List<uint> _extensions = [];
+    private readonly List<uint> _imports = [];
+    private readonly List<uint> _memoryModel = [];
+    private readonly List<uint> _entryPoints = [];
+    private readonly List<uint> _executionModes = [];
+    private readonly List<uint> _debug = [];
+    private readonly List<uint> _annotations = [];
+    private readonly List<uint> _typesConstantsGlobals = [];
+    private readonly List<uint> _functions = [];
+    private readonly Dictionary<(uint Width, bool Signed), uint> _integerTypes = [];
+    private readonly Dictionary<uint, uint> _floatTypes = [];
+    private readonly Dictionary<(uint Component, uint Count), uint> _vectorTypes = [];
+    private readonly Dictionary<
+        (
+            uint SampledType,
+            SpirvImageDim Dimension,
+            bool Depth,
+            bool Arrayed,
+            bool Multisampled,
+            uint Sampled,
+            SpirvImageFormat Format
+        ),
+        uint> _imageTypes = [];
+    private readonly Dictionary<uint, uint> _sampledImageTypes = [];
+    private readonly Dictionary<(SpirvStorageClass Storage, uint Type), uint> _pointerTypes = [];
+    private readonly Dictionary<(uint Element, uint Count), uint> _arrayTypes = [];
+    private readonly Dictionary<uint, uint> _runtimeArrayTypes = [];
+    private readonly Dictionary<string, uint> _functionTypes = [];
+    private readonly Dictionary<(uint Type, ulong Value), uint> _constants = [];
+    private readonly HashSet<SpirvCapability> _declaredCapabilities = [];
+    private readonly Dictionary<string, uint> _extInstImports = [];
+    private uint _nextId = 1;
+    private uint? _voidType;
+    private uint? _boolType;
+
+    public uint AllocateId() => _nextId++;
+
+    public void AddCapability(SpirvCapability capability)
+    {
+        if (_declaredCapabilities.Add(capability))
+        {
+            Emit(_capabilities, SpirvOp.Capability, (uint)capability);
+        }
+    }
+
+    public void AddExtension(string extension) =>
+        EmitWithString(_extensions, SpirvOp.Extension, [], extension);
+
+    public uint ImportExtInst(string name)
+    {
+        if (_extInstImports.TryGetValue(name, out var existing))
+        {
+            return existing;
+        }
+
+        var id = AllocateId();
+        EmitWithString(_imports, SpirvOp.ExtInstImport, [id], name);
+        _extInstImports.Add(name, id);
+        return id;
+    }
+
+    public void SetLogicalGlsl450MemoryModel() =>
+        Emit(_memoryModel, SpirvOp.MemoryModel, 0, 1);
+
+    public void AddEntryPoint(
+        SpirvExecutionModel model,
+        uint function,
+        string name,
+        IReadOnlyList<uint> interfaces)
+    {
+        var prefix = new uint[2 + interfaces.Count];
+        prefix[0] = (uint)model;
+        prefix[1] = function;
+        for (var index = 0; index < interfaces.Count; index++)
+        {
+            prefix[index + 2] = interfaces[index];
+        }
+
+        EmitWithString(_entryPoints, SpirvOp.EntryPoint, prefix, name, stringBeforeTailCount: 2);
+    }
+
+    public void AddExecutionMode(uint function, SpirvExecutionMode mode, params uint[] operands)
+    {
+        var values = new uint[2 + operands.Length];
+        values[0] = function;
+        values[1] = (uint)mode;
+        operands.CopyTo(values, 2);
+        Emit(_executionModes, SpirvOp.ExecutionMode, values);
+    }
+
+    public void AddName(uint target, string name) =>
+        EmitWithString(_debug, SpirvOp.Name, [target], name);
+
+    public void AddDecoration(uint target, SpirvDecoration decoration, params uint[] operands)
+    {
+        var values = new uint[2 + operands.Length];
+        values[0] = target;
+        values[1] = (uint)decoration;
+        operands.CopyTo(values, 2);
+        Emit(_annotations, SpirvOp.Decorate, values);
+    }
+
+    public void AddMemberDecoration(
+        uint target,
+        uint member,
+        SpirvDecoration decoration,
+        params uint[] operands)
+    {
+        var values = new uint[3 + operands.Length];
+        values[0] = target;
+        values[1] = member;
+        values[2] = (uint)decoration;
+        operands.CopyTo(values, 3);
+        EmitRaw(_annotations, 72, values);
+    }
+
+    public uint TypeVoid()
+    {
+        if (_voidType is { } existing)
+        {
+            return existing;
+        }
+
+        var id = AllocateId();
+        Emit(_typesConstantsGlobals, SpirvOp.TypeVoid, id);
+        _voidType = id;
+        return id;
+    }
+
+    public uint TypeBool()
+    {
+        if (_boolType is { } existing)
+        {
+            return existing;
+        }
+
+        var id = AllocateId();
+        Emit(_typesConstantsGlobals, SpirvOp.TypeBool, id);
+        _boolType = id;
+        return id;
+    }
+
+    public uint TypeInt(uint width, bool signed)
+    {
+        var key = (width, signed);
+        if (_integerTypes.TryGetValue(key, out var existing))
+        {
+            return existing;
+        }
+
+        var id = AllocateId();
+        Emit(_typesConstantsGlobals, SpirvOp.TypeInt, id, width, signed ? 1u : 0u);
+        _integerTypes.Add(key, id);
+        return id;
+    }
+
+    public uint TypeFloat(uint width)
+    {
+        if (_floatTypes.TryGetValue(width, out var existing))
+        {
+            return existing;
+        }
+
+        var id = AllocateId();
+        Emit(_typesConstantsGlobals, SpirvOp.TypeFloat, id, width);
+        _floatTypes.Add(width, id);
+        return id;
+    }
+
+    public uint TypeVector(uint componentType, uint count)
+    {
+        var key = (componentType, count);
+        if (_vectorTypes.TryGetValue(key, out var existing))
+        {
+            return existing;
+        }
+
+        var id = AllocateId();
+        Emit(_typesConstantsGlobals, SpirvOp.TypeVector, id, componentType, count);
+        _vectorTypes.Add(key, id);
+        return id;
+    }
+
+    public uint TypeImage(
+        uint sampledType,
+        SpirvImageDim dimension,
+        bool depth,
+        bool arrayed,
+        bool multisampled,
+        uint sampled,
+        SpirvImageFormat format)
+    {
+        var key = (
+            sampledType,
+            dimension,
+            depth,
+            arrayed,
+            multisampled,
+            sampled,
+            format);
+        if (_imageTypes.TryGetValue(key, out var existing))
+        {
+            return existing;
+        }
+
+        var id = AllocateId();
+        Emit(
+            _typesConstantsGlobals,
+            SpirvOp.TypeImage,
+            id,
+            sampledType,
+            (uint)dimension,
+            depth ? 1u : 0u,
+            arrayed ? 1u : 0u,
+            multisampled ? 1u : 0u,
+            sampled,
+            (uint)format);
+        _imageTypes.Add(key, id);
+        return id;
+    }
+
+    public uint TypeSampledImage(uint imageType)
+    {
+        if (_sampledImageTypes.TryGetValue(imageType, out var existing))
+        {
+            return existing;
+        }
+
+        var id = AllocateId();
+        Emit(_typesConstantsGlobals, SpirvOp.TypeSampledImage, id, imageType);
+        _sampledImageTypes.Add(imageType, id);
+        return id;
+    }
+
+    public uint TypeArray(uint elementType, uint count)
+    {
+        var key = (elementType, count);
+        if (_arrayTypes.TryGetValue(key, out var existing))
+        {
+            return existing;
+        }
+
+        var length = Constant(TypeInt(32, false), count);
+        var id = AllocateId();
+        Emit(_typesConstantsGlobals, SpirvOp.TypeArray, id, elementType, length);
+        _arrayTypes.Add(key, id);
+        return id;
+    }
+
+    public uint TypeRuntimeArray(uint elementType)
+    {
+        if (_runtimeArrayTypes.TryGetValue(elementType, out var existing))
+        {
+            return existing;
+        }
+
+        var id = AllocateId();
+        Emit(_typesConstantsGlobals, SpirvOp.TypeRuntimeArray, id, elementType);
+        _runtimeArrayTypes.Add(elementType, id);
+        return id;
+    }
+
+    public uint TypeStruct(params uint[] memberTypes)
+    {
+        var id = AllocateId();
+        var operands = new uint[memberTypes.Length + 1];
+        operands[0] = id;
+        memberTypes.CopyTo(operands, 1);
+        Emit(_typesConstantsGlobals, SpirvOp.TypeStruct, operands);
+        return id;
+    }
+
+    public uint TypePointer(SpirvStorageClass storageClass, uint type)
+    {
+        var key = (storageClass, type);
+        if (_pointerTypes.TryGetValue(key, out var existing))
+        {
+            return existing;
+        }
+
+        var id = AllocateId();
+        Emit(_typesConstantsGlobals, SpirvOp.TypePointer, id, (uint)storageClass, type);
+        _pointerTypes.Add(key, id);
+        return id;
+    }
+
+    public uint TypeFunction(uint returnType, params uint[] parameterTypes)
+    {
+        var key = returnType + ":" + string.Join(',', parameterTypes);
+        if (_functionTypes.TryGetValue(key, out var existing))
+        {
+            return existing;
+        }
+
+        var id = AllocateId();
+        var operands = new uint[parameterTypes.Length + 2];
+        operands[0] = id;
+        operands[1] = returnType;
+        parameterTypes.CopyTo(operands, 2);
+        Emit(_typesConstantsGlobals, SpirvOp.TypeFunction, operands);
+        _functionTypes.Add(key, id);
+        return id;
+    }
+
+    public uint ConstantBool(bool value)
+    {
+        var type = TypeBool();
+        var key = (type, value ? 1UL : 0UL);
+        if (_constants.TryGetValue(key, out var existing))
+        {
+            return existing;
+        }
+
+        var id = AllocateId();
+        Emit(
+            _typesConstantsGlobals,
+            value ? SpirvOp.ConstantTrue : SpirvOp.ConstantFalse,
+            type,
+            id);
+        _constants.Add(key, id);
+        return id;
+    }
+
+    public uint Constant(uint type, uint value)
+    {
+        var key = (type, (ulong)value);
+        if (_constants.TryGetValue(key, out var existing))
+        {
+            return existing;
+        }
+
+        var id = AllocateId();
+        Emit(_typesConstantsGlobals, SpirvOp.Constant, type, id, value);
+        _constants.Add(key, id);
+        return id;
+    }
+
+    public uint Constant64(uint type, ulong value)
+    {
+        var key = (type, value);
+        if (_constants.TryGetValue(key, out var existing))
+        {
+            return existing;
+        }
+
+        var id = AllocateId();
+        Emit(
+            _typesConstantsGlobals,
+            SpirvOp.Constant,
+            type,
+            id,
+            (uint)value,
+            (uint)(value >> 32));
+        _constants.Add(key, id);
+        return id;
+    }
+
+    public uint ConstantFloat(uint type, float value) =>
+        Constant(type, BitConverter.SingleToUInt32Bits(value));
+
+    public uint ConstantComposite(uint type, params uint[] constituents)
+    {
+        var id = AllocateId();
+        var operands = new uint[constituents.Length + 2];
+        operands[0] = type;
+        operands[1] = id;
+        constituents.CopyTo(operands, 2);
+        Emit(_typesConstantsGlobals, SpirvOp.ConstantComposite, operands);
+        return id;
+    }
+
+    public uint ConstantNull(uint type)
+    {
+        var key = (type, ulong.MaxValue);
+        if (_constants.TryGetValue(key, out var existing))
+        {
+            return existing;
+        }
+
+        var id = AllocateId();
+        Emit(_typesConstantsGlobals, SpirvOp.ConstantNull, type, id);
+        _constants.Add(key, id);
+        return id;
+    }
+
+    public uint AddGlobalVariable(
+        uint pointerType,
+        SpirvStorageClass storageClass,
+        uint? initializer = null)
+    {
+        var id = AllocateId();
+        if (initializer.HasValue)
+        {
+            Emit(
+                _typesConstantsGlobals,
+                SpirvOp.Variable,
+                pointerType,
+                id,
+                (uint)storageClass,
+                initializer.Value);
+        }
+        else
+        {
+            Emit(_typesConstantsGlobals, SpirvOp.Variable, pointerType, id, (uint)storageClass);
+        }
+
+        return id;
+    }
+
+    public uint BeginFunction(uint returnType, uint functionType)
+    {
+        var id = AllocateId();
+        Emit(_functions, SpirvOp.Function, returnType, id, 0, functionType);
+        return id;
+    }
+
+    public uint AddFunctionParameter(uint type) =>
+        EmitResult(_functions, SpirvOp.FunctionParameter, type);
+
+    public uint AddLabel(uint? id = null)
+    {
+        var result = id ?? AllocateId();
+        Emit(_functions, SpirvOp.Label, result);
+        return result;
+    }
+
+    public uint AddFunctionVariable(uint pointerType, uint? initializer = null)
+    {
+        var id = AllocateId();
+        if (initializer.HasValue)
+        {
+            Emit(
+                _functions,
+                SpirvOp.Variable,
+                pointerType,
+                id,
+                (uint)SpirvStorageClass.Function,
+                initializer.Value);
+        }
+        else
+        {
+            Emit(
+                _functions,
+                SpirvOp.Variable,
+                pointerType,
+                id,
+                (uint)SpirvStorageClass.Function);
+        }
+
+        return id;
+    }
+
+    public uint AddInstruction(SpirvOp opcode, uint resultType, params uint[] operands) =>
+        EmitResult(_functions, opcode, resultType, operands);
+
+    public void AddStatement(SpirvOp opcode, params uint[] operands) =>
+        Emit(_functions, opcode, operands);
+
+    public void EndFunction() => Emit(_functions, SpirvOp.FunctionEnd);
+
+    public byte[] Build()
+    {
+        if (_memoryModel.Count == 0)
+        {
+            SetLogicalGlsl450MemoryModel();
+        }
+
+        var wordCount =
+            5 +
+            _capabilities.Count +
+            _extensions.Count +
+            _imports.Count +
+            _memoryModel.Count +
+            _entryPoints.Count +
+            _executionModes.Count +
+            _debug.Count +
+            _annotations.Count +
+            _typesConstantsGlobals.Count +
+            _functions.Count;
+        var words = new uint[wordCount];
+        var offset = 0;
+        WriteWord(Magic);
+        WriteWord(Version15);
+        WriteWord(Generator);
+        WriteWord(_nextId);
+        WriteWord(0);
+        WriteSection(_capabilities);
+        WriteSection(_extensions);
+        WriteSection(_imports);
+        WriteSection(_memoryModel);
+        WriteSection(_entryPoints);
+        WriteSection(_executionModes);
+        WriteSection(_debug);
+        WriteSection(_annotations);
+        WriteSection(_typesConstantsGlobals);
+        WriteSection(_functions);
+        var bytes = new byte[wordCount * sizeof(uint)];
+        Buffer.BlockCopy(words, 0, bytes, 0, bytes.Length);
+        return bytes;
+
+        void WriteWord(uint value)
+        {
+            words[offset++] = value;
+        }
+
+        void WriteSection(List<uint> section)
+        {
+            foreach (var value in section)
+            {
+                WriteWord(value);
+            }
+        }
+    }
+
+    private uint EmitResult(
+        List<uint> section,
+        SpirvOp opcode,
+        uint resultType,
+        params uint[] operands)
+    {
+        var result = AllocateId();
+        var values = new uint[operands.Length + 2];
+        values[0] = resultType;
+        values[1] = result;
+        operands.CopyTo(values, 2);
+        Emit(section, opcode, values);
+        return result;
+    }
+
+    private static void Emit(List<uint> section, SpirvOp opcode, params uint[] operands) =>
+        EmitRaw(section, (ushort)opcode, operands);
+
+    private static void EmitRaw(List<uint> section, ushort opcode, params uint[] operands)
+    {
+        section.Add(((uint)(operands.Length + 1) << 16) | opcode);
+        section.AddRange(operands);
+    }
+
+    private static void EmitWithString(
+        List<uint> section,
+        SpirvOp opcode,
+        IReadOnlyList<uint> prefix,
+        string value,
+        int stringBeforeTailCount = -1)
+    {
+        var encoded = EncodeString(value);
+        if (stringBeforeTailCount < 0)
+        {
+            var operands = new uint[prefix.Count + encoded.Length];
+            for (var index = 0; index < prefix.Count; index++)
+            {
+                operands[index] = prefix[index];
+            }
+
+            encoded.CopyTo(operands, prefix.Count);
+            Emit(section, opcode, operands);
+            return;
+        }
+
+        var result = new uint[prefix.Count + encoded.Length];
+        for (var index = 0; index < stringBeforeTailCount; index++)
+        {
+            result[index] = prefix[index];
+        }
+
+        encoded.CopyTo(result, stringBeforeTailCount);
+        for (var index = stringBeforeTailCount; index < prefix.Count; index++)
+        {
+            result[index + encoded.Length] = prefix[index];
+        }
+
+        Emit(section, opcode, result);
+    }
+
+    private static uint[] EncodeString(string value)
+    {
+        var byteCount = Encoding.UTF8.GetByteCount(value) + 1;
+        var words = new uint[(byteCount + 3) / 4];
+        Encoding.UTF8.GetBytes(value, System.Runtime.InteropServices.MemoryMarshal.AsBytes(words.AsSpan()));
+        return words;
+    }
+}

src/SharpEmu.Libs/Ampr/AmprExports.cs

@@ -3,6 +3,7 @@
 
 using SharpEmu.HLE;
 using SharpEmu.Libs.Kernel;
+using System.Buffers;
 using System.Buffers.Binary;
 using System.Collections.Concurrent;
 
@@ -25,6 +26,9 @@ public static class AmprExports
     private static readonly ConcurrentDictionary<ulong, CommandBufferState> _commandBuffers = new();
     private static readonly bool _traceAmpr =
         string.Equals(Environment.GetEnvironmentVariable("SHARPEMU_LOG_AMPR"), "1", StringComparison.Ordinal);
+    private static readonly bool _traceAmprReads =
+        _traceAmpr ||
+        string.Equals(Environment.GetEnvironmentVariable("SHARPEMU_LOG_AMPR_READS"), "1", StringComparison.Ordinal);
 
     private sealed class CommandBufferState
     {
@@ -77,12 +81,7 @@ public static class AmprExports
             return (int)OrbisGen2Result.ORBIS_GEN2_OK;
         }
 
-        var buffer = 0UL;
-        var size = 0UL;
-        _ = ctx.TryReadUInt64(commandBuffer + CommandBufferDataOffset, out buffer);
-        _ = ctx.TryReadUInt64(commandBuffer + CommandBufferSizeOffset, out size);
-
-        if (!InitializeCommandBuffer(ctx, commandBuffer, buffer, size, aux0, aux1, clear: false))
+        if (!InitializeCommandBuffer(ctx, commandBuffer, buffer: 0, size: 0, aux0, aux1, clear: false))
         {
             return (int)OrbisGen2Result.ORBIS_GEN2_ERROR_MEMORY_FAULT;
         }
@@ -106,8 +105,9 @@ public static class AmprExports
             return (int)OrbisGen2Result.ORBIS_GEN2_OK;
         }
 
-        if (!ctx.TryWriteUInt64(commandBuffer + CommandBufferAux0Offset, 0) ||
-            !ctx.TryWriteUInt64(commandBuffer + CommandBufferAux1Offset, 0))
+        Span<byte> auxiliaryPointers = stackalloc byte[sizeof(ulong) * 2];
+        auxiliaryPointers.Clear();
+        if (!ctx.Memory.TryWrite(commandBuffer + CommandBufferAux0Offset, auxiliaryPointers))
         {
             return (int)OrbisGen2Result.ORBIS_GEN2_ERROR_MEMORY_FAULT;
         }
@@ -181,8 +181,15 @@ public static class AmprExports
             return (int)OrbisGen2Result.ORBIS_GEN2_ERROR_INVALID_ARGUMENT;
         }
 
-        if (!ctx.TryReadUInt64(commandBuffer + CommandBufferDataOffset, out var buffer) ||
-            !ctx.TryReadUInt64(commandBuffer + CommandBufferSizeOffset, out var size) ||
+        Span<byte> bufferPointers = stackalloc byte[sizeof(ulong) * 2];
+        if (!ctx.Memory.TryRead(commandBuffer + CommandBufferDataOffset, bufferPointers))
+        {
+            return (int)OrbisGen2Result.ORBIS_GEN2_ERROR_MEMORY_FAULT;
+        }
+
+        var buffer = BinaryPrimitives.ReadUInt64LittleEndian(bufferPointers);
+        var size = BinaryPrimitives.ReadUInt64LittleEndian(bufferPointers[sizeof(ulong)..]);
+        if (
             !WriteCommandBufferPointers(ctx, commandBuffer, buffer, size, writeOffset: 0))
         {
             return (int)OrbisGen2Result.ORBIS_GEN2_ERROR_MEMORY_FAULT;
@@ -476,20 +483,34 @@ public static class AmprExports
         ulong aux1,
         bool clear)
     {
+        Span<byte> header = stackalloc byte[CommandBufferHeaderSize];
         if (clear)
         {
-            Span<byte> header = stackalloc byte[CommandBufferHeaderSize];
             header.Clear();
-            if (!ctx.Memory.TryWrite(commandBuffer, header))
+        }
+        else
+        {
+            if (!ctx.Memory.TryRead(commandBuffer, header))
             {
                 return false;
             }
+
+            buffer = BinaryPrimitives.ReadUInt64LittleEndian(header[(int)CommandBufferDataOffset..]);
+            size = BinaryPrimitives.ReadUInt64LittleEndian(header[(int)CommandBufferSizeOffset..]);
         }
 
-        return ctx.TryWriteUInt64(commandBuffer + CommandBufferSelfOffset, commandBuffer) &&
-               ctx.TryWriteUInt64(commandBuffer + CommandBufferAux0Offset, aux0) &&
-               ctx.TryWriteUInt64(commandBuffer + CommandBufferAux1Offset, aux1) &&
-               WriteCommandBufferPointers(ctx, commandBuffer, buffer, size, writeOffset: 0);
+        BinaryPrimitives.WriteUInt64LittleEndian(header[(int)CommandBufferSelfOffset..], commandBuffer);
+        BinaryPrimitives.WriteUInt64LittleEndian(header[(int)CommandBufferDataOffset..], buffer);
+        BinaryPrimitives.WriteUInt64LittleEndian(header[(int)CommandBufferSizeOffset..], size);
+        BinaryPrimitives.WriteUInt64LittleEndian(header[(int)CommandBufferAux0Offset..], aux0);
+        BinaryPrimitives.WriteUInt64LittleEndian(header[(int)CommandBufferAux1Offset..], aux1);
+        if (!ctx.Memory.TryWrite(commandBuffer, header))
+        {
+            return false;
+        }
+
+        UpdateCommandBufferState(commandBuffer, buffer, size, writeOffset: 0);
+        return true;
     }
 
     private static bool WriteCommandBufferPointers(CpuContext ctx, ulong commandBuffer, ulong buffer, ulong size)
@@ -504,6 +525,26 @@ public static class AmprExports
             return false;
         }
 
+        UpdateCommandBufferState(commandBuffer, buffer, size, writeOffset);
+
+        return true;
+    }
+
+    private static bool WriteVisibleCommandBufferPointers(CpuContext ctx, ulong commandBuffer, ulong buffer, ulong size)
+    {
+        Span<byte> pointers = stackalloc byte[sizeof(ulong) * 3];
+        BinaryPrimitives.WriteUInt64LittleEndian(pointers, commandBuffer);
+        BinaryPrimitives.WriteUInt64LittleEndian(pointers[sizeof(ulong)..], buffer);
+        BinaryPrimitives.WriteUInt64LittleEndian(pointers[(sizeof(ulong) * 2)..], size);
+        return ctx.Memory.TryWrite(commandBuffer + CommandBufferSelfOffset, pointers);
+    }
+
+    private static void UpdateCommandBufferState(
+        ulong commandBuffer,
+        ulong buffer,
+        ulong size,
+        ulong writeOffset)
+    {
         var state = _commandBuffers.GetOrAdd(commandBuffer, static _ => new CommandBufferState());
         lock (state)
         {
@@ -511,15 +552,6 @@ public static class AmprExports
             state.Size = size;
             state.WriteOffset = writeOffset;
         }
-
-        return true;
-    }
-
-    private static bool WriteVisibleCommandBufferPointers(CpuContext ctx, ulong commandBuffer, ulong buffer, ulong size)
-    {
-        return ctx.TryWriteUInt64(commandBuffer + CommandBufferSelfOffset, commandBuffer) &&
-               ctx.TryWriteUInt64(commandBuffer + CommandBufferDataOffset, buffer) &&
-               ctx.TryWriteUInt64(commandBuffer + CommandBufferSizeOffset, size);
     }
 
     private static bool TryGetCommandBufferState(
@@ -540,9 +572,11 @@ public static class AmprExports
             return true;
         }
 
-        if (ctx.TryReadUInt64(commandBuffer + CommandBufferDataOffset, out buffer) &&
-            ctx.TryReadUInt64(commandBuffer + CommandBufferSizeOffset, out size))
+        Span<byte> pointers = stackalloc byte[sizeof(ulong) * 2];
+        if (ctx.Memory.TryRead(commandBuffer + CommandBufferDataOffset, pointers))
         {
+            buffer = BinaryPrimitives.ReadUInt64LittleEndian(pointers);
+            size = BinaryPrimitives.ReadUInt64LittleEndian(pointers[sizeof(ulong)..]);
             state = _commandBuffers.GetOrAdd(commandBuffer, static _ => new CommandBufferState());
             lock (state)
             {
@@ -595,12 +629,18 @@ public static class AmprExports
             return (int)OrbisGen2Result.ORBIS_GEN2_ERROR_INVALID_ARGUMENT;
         }
 
-        const int ChunkSize = 64 * 1024;
-        var buffer = new byte[(int)Math.Min((ulong)ChunkSize, size)];
+        const int ChunkSize = 1024 * 1024;
+        var buffer = ArrayPool<byte>.Shared.Rent((int)Math.Min((ulong)ChunkSize, size));
 
         try
         {
-            using var stream = new FileStream(hostPath, FileMode.Open, FileAccess.Read, FileShare.ReadWrite | FileShare.Delete);
+            using var stream = new FileStream(
+                hostPath,
+                FileMode.Open,
+                FileAccess.Read,
+                FileShare.ReadWrite | FileShare.Delete,
+                ChunkSize,
+                FileOptions.SequentialScan);
             if (fileOffset >= (ulong)stream.Length)
             {
                 return (int)OrbisGen2Result.ORBIS_GEN2_OK;
@@ -633,6 +673,10 @@ public static class AmprExports
         {
             return (int)OrbisGen2Result.ORBIS_GEN2_ERROR_NOT_FOUND;
         }
+        finally
+        {
+            ArrayPool<byte>.Shared.Return(buffer);
+        }
 
         return (int)OrbisGen2Result.ORBIS_GEN2_OK;
     }
@@ -801,7 +845,7 @@ public static class AmprExports
         string? hostPath,
         int result)
     {
-        if (!_traceAmpr)
+        if (!_traceAmprReads)
         {
             return;
         }

src/SharpEmu.Libs/Kernel/KernelExports.cs

@@ -204,6 +204,16 @@ public static class KernelExports
         var nameAddress = ctx[CpuRegister.R8];
         var name = nameAddress == 0 ? string.Empty : ReadCString(ctx, nameAddress, 256);
         var threadHandle = KernelPthreadState.CreateThreadHandle(name);
+        KernelPthreadExtendedCompatExports.GetThreadStartScheduling(
+            ctx,
+            attrAddress,
+            out var priority,
+            out var affinityMask);
+        KernelPthreadExtendedCompatExports.RegisterThreadStart(
+            threadHandle,
+            name,
+            priority,
+            affinityMask);
         if (threadIdAddress != 0 && !ctx.TryWriteUInt64(threadIdAddress, threadHandle))
         {
             return (int)OrbisGen2Result.ORBIS_GEN2_ERROR_MEMORY_FAULT;
@@ -212,13 +222,22 @@ public static class KernelExports
         if (ShouldTracePthread())
         {
             Console.Error.WriteLine(
-                $"[LOADER][TRACE] pthread_create: out=0x{threadIdAddress:X16} attr=0x{attrAddress:X16} entry=0x{entryAddress:X16} arg=0x{argument:X16} name_ptr=0x{nameAddress:X16} name='{name}' -> thread=0x{threadHandle:X16}");
+                $"[LOADER][TRACE] pthread_create: out=0x{threadIdAddress:X16} attr=0x{attrAddress:X16} " +
+                $"entry=0x{entryAddress:X16} arg=0x{argument:X16} name_ptr=0x{nameAddress:X16} " +
+                $"name='{name}' priority={priority} affinity=0x{affinityMask:X} -> thread=0x{threadHandle:X16}");
         }
 
         var scheduler = GuestThreadExecution.Scheduler;
         if (scheduler is not null && entryAddress != 0)
         {
-            var request = new GuestThreadStartRequest(threadHandle, entryAddress, argument, attrAddress, name);
+            var request = new GuestThreadStartRequest(
+                threadHandle,
+                entryAddress,
+                argument,
+                attrAddress,
+                name,
+                priority,
+                affinityMask);
             if (!scheduler.TryStartThread(ctx, request, out var error))
             {
                 Console.Error.WriteLine(
@@ -287,10 +306,6 @@ public static class KernelExports
     {
         var threadId = ctx[CpuRegister.Rdi];
         var returnValueAddress = ctx[CpuRegister.Rsi];
-        if (returnValueAddress != 0 && !ctx.TryWriteUInt64(returnValueAddress, 0))
-        {
-            return (int)OrbisGen2Result.ORBIS_GEN2_ERROR_MEMORY_FAULT;
-        }
 
         if (ShouldTracePthread())
         {
@@ -298,6 +313,30 @@ public static class KernelExports
                 $"[LOADER][TRACE] pthread_join: thread=0x{threadId:X16} retval_out=0x{returnValueAddress:X16}");
         }
 
+        var returnValue = 0UL;
+        if (GuestThreadExecution.Scheduler is { } scheduler &&
+            !scheduler.TryJoinThread(ctx, threadId, out returnValue, out var error))
+        {
+            Console.Error.WriteLine(
+                $"[LOADER][ERROR] pthread_join: thread=0x{threadId:X16}: {error}");
+            var result = string.Equals(
+                error,
+                "thread cannot join itself",
+                StringComparison.Ordinal)
+                ? OrbisGen2Result.ORBIS_GEN2_ERROR_INVALID_ARGUMENT
+                : OrbisGen2Result.ORBIS_GEN2_ERROR_NOT_FOUND;
+            ctx[CpuRegister.Rax] = unchecked((ulong)(int)result);
+            return (int)result;
+        }
+
+        if (returnValueAddress != 0 &&
+            !ctx.TryWriteUInt64(returnValueAddress, returnValue))
+        {
+            ctx[CpuRegister.Rax] =
+                unchecked((ulong)(int)OrbisGen2Result.ORBIS_GEN2_ERROR_MEMORY_FAULT);
+            return (int)OrbisGen2Result.ORBIS_GEN2_ERROR_MEMORY_FAULT;
+        }
+
         ctx[CpuRegister.Rax] = 0;
         return (int)OrbisGen2Result.ORBIS_GEN2_OK;
     }

src/SharpEmu.Libs/Kernel/KernelMemoryCompatExports.cs

@@ -99,10 +99,12 @@ public static class KernelMemoryCompatExports
     private static readonly object _tlsGate = new();
     private static readonly object _ioTraceGate = new();
     private static readonly object _statCacheGate = new();
+    private static readonly object _guestMountGate = new();
     private static readonly Dictionary<ulong, DirectAllocation> _directAllocations = new();
     private static readonly Dictionary<ulong, LibcHeapAllocation> _libcAllocations = new();
     private static readonly Dictionary<ulong, MappedRegion> _mappedRegions = new();
     private static readonly Dictionary<ulong, ulong> _tlsModuleBlocks = new();
+    private static readonly Dictionary<string, string> _guestMounts = new(StringComparer.OrdinalIgnoreCase);
     private static readonly HashSet<string> _tracedStatResults = new(StringComparer.Ordinal);
     private static readonly HashSet<string> _negativeStatCache = new(StringComparer.OrdinalIgnoreCase);
     private static long _nextFileDescriptor = 2;
@@ -153,6 +155,32 @@ public static class KernelMemoryCompatExports
     private readonly record struct MappedRegion(ulong Address, ulong Length, int Protection, bool IsFlexible, bool IsDirect, ulong DirectStart);
     private readonly record struct BatchMapEntry(ulong Start, ulong Offset, ulong Length, byte Protection, byte Type, int Operation);
 
+    public static void RegisterGuestPathMount(string guestMountPoint, string hostRoot)
+    {
+        ArgumentException.ThrowIfNullOrWhiteSpace(guestMountPoint);
+        ArgumentException.ThrowIfNullOrWhiteSpace(hostRoot);
+
+        var normalizedMountPoint = NormalizeGuestStatCachePath(guestMountPoint);
+        if (normalizedMountPoint is null || normalizedMountPoint == "/")
+        {
+            throw new ArgumentException("Guest mount point must name a directory.", nameof(guestMountPoint));
+        }
+
+        var normalizedHostRoot = Path.GetFullPath(hostRoot);
+        Directory.CreateDirectory(normalizedHostRoot);
+        lock (_guestMountGate)
+        {
+            _guestMounts[normalizedMountPoint] = normalizedHostRoot;
+        }
+
+        lock (_statCacheGate)
+        {
+            _negativeStatCache.RemoveWhere(path =>
+                string.Equals(path, normalizedMountPoint, StringComparison.OrdinalIgnoreCase) ||
+                path.StartsWith(normalizedMountPoint + "/", StringComparison.OrdinalIgnoreCase));
+        }
+    }
+
     internal static bool TryAllocateHleData(
         CpuContext ctx,
         ulong length,
@@ -602,9 +630,7 @@ public static class KernelMemoryCompatExports
         }
         else
         {
-            ulong NextGpArg() => vaCursor.NextGpArg();
-            double NextFloatArg() => vaCursor.NextFloatArg();
-            rendered = FormatString(ctx, format, NextGpArg, NextFloatArg);
+            rendered = FormatString(ctx, format, ref vaCursor);
         }
 
         Console.Write(rendered);
@@ -2908,9 +2934,8 @@ public static class KernelMemoryCompatExports
 
         var format = Encoding.UTF8.GetString(formatBytes);
         var result = FormatString(ctx, format);
-        var outputBytes = Encoding.UTF8.GetBytes(result);
 
-        return WriteSnprintfOutput(ctx, destination, bufferSize, outputBytes);
+        return WriteSnprintfOutput(ctx, destination, bufferSize, result);
     }
 
     private static int VsnprintfCore(CpuContext ctx)
@@ -2931,12 +2956,9 @@ public static class KernelMemoryCompatExports
             return WriteSnprintfOutput(ctx, destination, bufferSize, formatBytes);
         }
 
-        ulong NextGpArg() => vaCursor.NextGpArg();
-        double NextFloatArg() => vaCursor.NextFloatArg();
-        var rendered = FormatString(ctx, format, NextGpArg, NextFloatArg);
+        var rendered = FormatString(ctx, format, ref vaCursor);
 
-        var outputBytes = Encoding.UTF8.GetBytes(rendered);
-        return WriteSnprintfOutput(ctx, destination, bufferSize, outputBytes);
+        return WriteSnprintfOutput(ctx, destination, bufferSize, rendered);
     }
 
     private static int SwprintfCore(CpuContext ctx)
@@ -2977,9 +2999,7 @@ public static class KernelMemoryCompatExports
         else
         {
             TraceWidePrintfVaList(ctx, "vswprintf", format, vaListAddress, vaCursor);
-            ulong NextGpArg() => vaCursor.NextGpArg();
-            double NextFloatArg() => vaCursor.NextFloatArg();
-            rendered = FormatString(ctx, format, NextGpArg, NextFloatArg);
+            rendered = FormatString(ctx, format, ref vaCursor);
         }
 
         TraceWidePrintf(ctx, "vswprintf", destination, bufferSize, format, rendered);
@@ -2994,10 +3014,8 @@ public static class KernelMemoryCompatExports
             return false;
         }
 
-        if (!TryReadUInt32Compat(ctx, vaListAddress + 0, out var gpOffset) ||
-            !TryReadUInt32Compat(ctx, vaListAddress + 4, out var fpOffset) ||
-            !TryReadUInt64Compat(ctx, vaListAddress + 8, out var overflowArgArea) ||
-            !TryReadUInt64Compat(ctx, vaListAddress + 16, out var regSaveArea))
+        Span<byte> vaList = stackalloc byte[24];
+        if (!TryReadCompat(ctx, vaListAddress, vaList))
         {
             return false;
         }
@@ -3005,10 +3023,10 @@ public static class KernelMemoryCompatExports
         cursor = new SysVAmd64VaListCursor(
             ctx,
             vaListAddress,
-            gpOffset,
-            fpOffset,
-            overflowArgArea,
-            regSaveArea);
+            BinaryPrimitives.ReadUInt32LittleEndian(vaList),
+            BinaryPrimitives.ReadUInt32LittleEndian(vaList[4..]),
+            BinaryPrimitives.ReadUInt64LittleEndian(vaList[8..]),
+            BinaryPrimitives.ReadUInt64LittleEndian(vaList[16..]));
         return true;
     }
 
@@ -3038,6 +3056,21 @@ public static class KernelMemoryCompatExports
         return (int)OrbisGen2Result.ORBIS_GEN2_OK;
     }
 
+    private static int WriteSnprintfOutput(
+        CpuContext ctx,
+        ulong destination,
+        ulong bufferSize,
+        string output)
+    {
+        if (bufferSize == 0 || destination == 0)
+        {
+            ctx[CpuRegister.Rax] = unchecked((ulong)Encoding.UTF8.GetByteCount(output));
+            return (int)OrbisGen2Result.ORBIS_GEN2_OK;
+        }
+
+        return WriteSnprintfOutput(ctx, destination, bufferSize, Encoding.UTF8.GetBytes(output));
+    }
+
     private static int WriteSwprintfOutput(
         CpuContext ctx,
         ulong destination,
@@ -3329,30 +3362,8 @@ public static class KernelMemoryCompatExports
 
     internal static string FormatStringFromVarArgs(CpuContext ctx, string format, int firstGpArgIndex)
     {
-        var gpIndex = Math.Max(0, firstGpArgIndex);
-
-        ulong GetGpArg(int index)
-        {
-            return index switch
-            {
-                0 => ctx[CpuRegister.Rdi],
-                1 => ctx[CpuRegister.Rsi],
-                2 => ctx[CpuRegister.Rdx],
-                3 => ctx[CpuRegister.Rcx],
-                4 => ctx[CpuRegister.R8],
-                5 => ctx[CpuRegister.R9],
-                _ => ReadStackArg(ctx, (ulong)(index - 6) * 8)
-            };
-        }
-
-        ulong NextGpArg() => GetGpArg(gpIndex++);
-        double NextFloatArg()
-        {
-            var rawBits = NextGpArg();
-            return BitConverter.Int64BitsToDouble(unchecked((long)rawBits));
-        }
-
-        return FormatString(ctx, format, NextGpArg, NextFloatArg);
+        var argumentSource = new RegisterPrintfArgumentSource(ctx, Math.Max(0, firstGpArgIndex));
+        return FormatString(ctx, format, ref argumentSource);
     }
 
     private static string FormatString(CpuContext ctx, string format)
@@ -3360,13 +3371,13 @@ public static class KernelMemoryCompatExports
         return FormatStringFromVarArgs(ctx, format, firstGpArgIndex: 3);
     }
 
-    private static string FormatString(
+    private static string FormatString<TArgumentSource>(
         CpuContext ctx,
         string format,
-        Func<ulong> nextGpArg,
-        Func<double> nextFloatArg)
+        ref TArgumentSource argumentSource)
+        where TArgumentSource : struct, IPrintfArgumentSource
     {
-        var sb = new StringBuilder();
+        var sb = new StringBuilder(format.Length + 32);
 
         for (var i = 0; i < format.Length; i++)
         {
@@ -3405,7 +3416,7 @@ public static class KernelMemoryCompatExports
             var width = 0;
             if (i < format.Length && format[i] == '*')
             {
-                width = unchecked((int)nextGpArg());
+                width = unchecked((int)argumentSource.NextGpArg());
                 i++;
                 if (width < 0)
                 {
@@ -3428,7 +3439,7 @@ public static class KernelMemoryCompatExports
                 i++;
                 if (i < format.Length && format[i] == '*')
                 {
-                    precision = unchecked((int)nextGpArg());
+                    precision = unchecked((int)argumentSource.NextGpArg());
                     i++;
                 }
                 else if (i < format.Length && char.IsDigit(format[i]))
@@ -3482,14 +3493,14 @@ public static class KernelMemoryCompatExports
                     {
                         long value = lengthMod switch
                         {
-                            "hh" => unchecked((sbyte)nextGpArg()),
-                            "h" => unchecked((short)nextGpArg()),
-                            "l" => unchecked((long)nextGpArg()),
-                            "ll" => unchecked((long)nextGpArg()),
-                            "j" => unchecked((long)nextGpArg()),
-                            "z" => unchecked((long)nextGpArg()),
-                            "t" => unchecked((long)nextGpArg()),
-                            _ => unchecked((int)nextGpArg())
+                            "hh" => unchecked((sbyte)argumentSource.NextGpArg()),
+                            "h" => unchecked((short)argumentSource.NextGpArg()),
+                            "l" => unchecked((long)argumentSource.NextGpArg()),
+                            "ll" => unchecked((long)argumentSource.NextGpArg()),
+                            "j" => unchecked((long)argumentSource.NextGpArg()),
+                            "z" => unchecked((long)argumentSource.NextGpArg()),
+                            "t" => unchecked((long)argumentSource.NextGpArg()),
+                            _ => unchecked((int)argumentSource.NextGpArg())
                         };
 
                         var formatted = value.ToString();
@@ -3506,14 +3517,14 @@ public static class KernelMemoryCompatExports
                     {
                         ulong value = lengthMod switch
                         {
-                            "hh" => (byte)nextGpArg(),
-                            "h" => (ushort)nextGpArg(),
-                            "l" => nextGpArg(),
-                            "ll" => nextGpArg(),
-                            "j" => nextGpArg(),
-                            "z" => nextGpArg(),
-                            "t" => nextGpArg(),
-                            _ => (uint)nextGpArg()
+                            "hh" => (byte)argumentSource.NextGpArg(),
+                            "h" => (ushort)argumentSource.NextGpArg(),
+                            "l" => argumentSource.NextGpArg(),
+                            "ll" => argumentSource.NextGpArg(),
+                            "j" => argumentSource.NextGpArg(),
+                            "z" => argumentSource.NextGpArg(),
+                            "t" => argumentSource.NextGpArg(),
+                            _ => (uint)argumentSource.NextGpArg()
                         };
 
                         var formatted = value.ToString();
@@ -3526,14 +3537,14 @@ public static class KernelMemoryCompatExports
                     {
                         ulong value = lengthMod switch
                         {
-                            "hh" => (byte)nextGpArg(),
-                            "h" => (ushort)nextGpArg(),
-                            "l" => nextGpArg(),
-                            "ll" => nextGpArg(),
-                            "j" => nextGpArg(),
-                            "z" => nextGpArg(),
-                            "t" => nextGpArg(),
-                            _ => (uint)nextGpArg()
+                            "hh" => (byte)argumentSource.NextGpArg(),
+                            "h" => (ushort)argumentSource.NextGpArg(),
+                            "l" => argumentSource.NextGpArg(),
+                            "ll" => argumentSource.NextGpArg(),
+                            "j" => argumentSource.NextGpArg(),
+                            "z" => argumentSource.NextGpArg(),
+                            "t" => argumentSource.NextGpArg(),
+                            _ => (uint)argumentSource.NextGpArg()
                         };
 
                         var formatted = specifier == 'x'
@@ -3551,14 +3562,14 @@ public static class KernelMemoryCompatExports
                     {
                         ulong value = lengthMod switch
                         {
-                            "hh" => (byte)nextGpArg(),
-                            "h" => (ushort)nextGpArg(),
-                            "l" => nextGpArg(),
-                            "ll" => nextGpArg(),
-                            "j" => nextGpArg(),
-                            "z" => nextGpArg(),
-                            "t" => nextGpArg(),
-                            _ => (uint)nextGpArg()
+                            "hh" => (byte)argumentSource.NextGpArg(),
+                            "h" => (ushort)argumentSource.NextGpArg(),
+                            "l" => argumentSource.NextGpArg(),
+                            "ll" => argumentSource.NextGpArg(),
+                            "j" => argumentSource.NextGpArg(),
+                            "z" => argumentSource.NextGpArg(),
+                            "t" => argumentSource.NextGpArg(),
+                            _ => (uint)argumentSource.NextGpArg()
                         };
 
                         var formatted = Convert.ToString((long)value, 8);
@@ -3571,7 +3582,7 @@ public static class KernelMemoryCompatExports
 
                 case 'p':
                     {
-                        var value = nextGpArg();
+                        var value = argumentSource.NextGpArg();
                         var formatted = value == 0
                             ? "(nil)"
                             : $"0x{value:X}";
@@ -3581,7 +3592,7 @@ public static class KernelMemoryCompatExports
 
                 case 's':
                     {
-                        var strAddr = nextGpArg();
+                        var strAddr = argumentSource.NextGpArg();
                         TracePrintfStringArgument(ctx, lengthMod, strAddr);
                         if (strAddr == 0)
                         {
@@ -3620,14 +3631,14 @@ public static class KernelMemoryCompatExports
                         string renderedChar;
                         if (lengthMod == "l")
                         {
-                            var scalar = unchecked((ushort)nextGpArg());
+                            var scalar = unchecked((ushort)argumentSource.NextGpArg());
                             renderedChar = TryConvertWideScalarToString(scalar, out var wideCharText)
                                 ? wideCharText
                                 : "?";
                         }
                         else
                         {
-                            renderedChar = ((char)(byte)nextGpArg()).ToString();
+                            renderedChar = ((char)(byte)argumentSource.NextGpArg()).ToString();
                         }
 
                         sb.Append(PadString(renderedChar, width, leftAlign, false));
@@ -3641,7 +3652,7 @@ public static class KernelMemoryCompatExports
                 case 'g':
                 case 'G':
                     {
-                        var value = nextFloatArg();
+                        var value = argumentSource.NextFloatArg();
 
                         var formatStr = precision >= 0
                             ? $"{{0:{specifier}{precision}}}"
@@ -3659,7 +3670,7 @@ public static class KernelMemoryCompatExports
 
                 case 'n':
                     {
-                        var addr = nextGpArg();
+                        var addr = argumentSource.NextGpArg();
                         if (addr != 0)
                         {
                             _ = TryWriteInt32(ctx, addr, sb.Length);
@@ -3699,7 +3710,46 @@ public static class KernelMemoryCompatExports
         return leftAlign ? str + padding : padding + str;
     }
 
-    private struct SysVAmd64VaListCursor
+    private interface IPrintfArgumentSource
+    {
+        ulong NextGpArg();
+
+        double NextFloatArg();
+    }
+
+    private struct RegisterPrintfArgumentSource : IPrintfArgumentSource
+    {
+        private readonly CpuContext _ctx;
+        private int _gpIndex;
+
+        public RegisterPrintfArgumentSource(CpuContext ctx, int gpIndex)
+        {
+            _ctx = ctx;
+            _gpIndex = gpIndex;
+        }
+
+        public ulong NextGpArg()
+        {
+            var index = _gpIndex++;
+            return index switch
+            {
+                0 => _ctx[CpuRegister.Rdi],
+                1 => _ctx[CpuRegister.Rsi],
+                2 => _ctx[CpuRegister.Rdx],
+                3 => _ctx[CpuRegister.Rcx],
+                4 => _ctx[CpuRegister.R8],
+                5 => _ctx[CpuRegister.R9],
+                _ => ReadStackArg(_ctx, (ulong)(index - 6) * 8)
+            };
+        }
+
+        public double NextFloatArg()
+        {
+            return BitConverter.Int64BitsToDouble(unchecked((long)NextGpArg()));
+        }
+    }
+
+    private struct SysVAmd64VaListCursor : IPrintfArgumentSource
     {
         private const uint GpSaveAreaLimit = 48;
         private const uint FpSaveAreaLimit = 176;
@@ -4011,6 +4061,11 @@ public static class KernelMemoryCompatExports
             return guestPath;
         }
 
+        if (TryResolveRegisteredGuestMount(guestPath, out var mountedPath))
+        {
+            return mountedPath;
+        }
+
         if (guestPath.StartsWith("/devlog/app/", StringComparison.OrdinalIgnoreCase))
         {
             var relative = NormalizeMountRelativePath(guestPath["/devlog/app/".Length..]);
@@ -4109,6 +4164,51 @@ public static class KernelMemoryCompatExports
         return guestPath;
     }
 
+    private static bool TryResolveRegisteredGuestMount(string guestPath, out string hostPath)
+    {
+        hostPath = string.Empty;
+        var normalizedGuestPath = NormalizeGuestStatCachePath(guestPath);
+        if (normalizedGuestPath is null)
+        {
+            return false;
+        }
+
+        string? matchedMountPoint = null;
+        string? matchedHostRoot = null;
+        lock (_guestMountGate)
+        {
+            foreach (var (mountPoint, hostRoot) in _guestMounts)
+            {
+                if ((string.Equals(normalizedGuestPath, mountPoint, StringComparison.OrdinalIgnoreCase) ||
+                     normalizedGuestPath.StartsWith(mountPoint + "/", StringComparison.OrdinalIgnoreCase)) &&
+                    (matchedMountPoint is null || mountPoint.Length > matchedMountPoint.Length))
+                {
+                    matchedMountPoint = mountPoint;
+                    matchedHostRoot = hostRoot;
+                }
+            }
+        }
+
+        if (matchedMountPoint is null || matchedHostRoot is null)
+        {
+            return false;
+        }
+
+        var relativePath = normalizedGuestPath[matchedMountPoint.Length..].TrimStart('/');
+        var candidate = Path.GetFullPath(Path.Combine(
+            matchedHostRoot,
+            NormalizeMountRelativePath(relativePath)));
+        var rootWithSeparator = Path.TrimEndingDirectorySeparator(matchedHostRoot) + Path.DirectorySeparatorChar;
+        if (!string.Equals(candidate, matchedHostRoot, StringComparison.OrdinalIgnoreCase) &&
+            !candidate.StartsWith(rootWithSeparator, StringComparison.OrdinalIgnoreCase))
+        {
+            return false;
+        }
+
+        hostPath = candidate;
+        return true;
+    }
+
     private static string? ResolveApp0Root()
     {
         var cached = Volatile.Read(ref _cachedApp0Root);
@@ -4297,9 +4397,32 @@ public static class KernelMemoryCompatExports
         }
 
         const int maxChunkSize = 4096;
+        const int inlineChunkSize = 256;
+        Span<byte> inlineChunk = stackalloc byte[inlineChunkSize];
+        var firstPageRemaining = maxChunkSize - (int)(address & (maxChunkSize - 1));
+        var firstReadLength = Math.Min(limit, Math.Min(inlineChunkSize, firstPageRemaining));
+        var firstSpan = inlineChunk[..firstReadLength];
+        ulong offset = 0;
+        if (TryReadCompat(ctx, address, firstSpan))
+        {
+            var nulIndex = firstSpan.IndexOf((byte)0);
+            if (nulIndex >= 0)
+            {
+                bytes = firstSpan[..nulIndex].ToArray();
+                return true;
+            }
+
+            offset = unchecked((ulong)firstReadLength);
+        }
+
         var chunk = GC.AllocateUninitializedArray<byte>(Math.Min(maxChunkSize, limit));
         var writer = new ArrayBufferWriter<byte>(Math.Min(limit, 256));
-        ulong offset = 0;
+        if (offset != 0)
+        {
+            firstSpan.CopyTo(writer.GetSpan(firstReadLength));
+            writer.Advance(firstReadLength);
+        }
+
         while (offset < (ulong)limit)
         {
             var current = address + offset;
@@ -5265,6 +5388,37 @@ public static class KernelMemoryCompatExports
         }
     }
 
+    internal static bool TryReadTrackedLibcHeap(
+        ulong address,
+        Span<byte> destination)
+    {
+        if (destination.IsEmpty)
+        {
+            return true;
+        }
+
+        var length = (ulong)destination.Length;
+        lock (_libcAllocGate)
+        {
+            foreach (var (allocationAddress, allocation) in _libcAllocations)
+            {
+                var allocationSize = (ulong)allocation.Size;
+                var offset = address >= allocationAddress
+                    ? address - allocationAddress
+                    : ulong.MaxValue;
+                if (offset > allocationSize ||
+                    length > allocationSize - offset)
+                {
+                    continue;
+                }
+
+                return TryReadHostMemory(address, destination);
+            }
+        }
+
+        return false;
+    }
+
     private static bool TryAllocateLibcHeap(ulong requestedSize, nuint alignment, bool zeroFill, out ulong address)
     {
         address = 0;

src/SharpEmu.Libs/Kernel/KernelPthreadCompatExports.cs

@@ -11,7 +11,6 @@ namespace SharpEmu.Libs.Kernel;
 
 public static class KernelPthreadCompatExports
 {
-    private const int MutexTypeDefault = 0;
     private const int MutexTypeErrorCheck = 1;
     private const int MutexTypeRecursive = 2;
     private const int MutexTypeNormal = 3;
@@ -44,7 +43,7 @@ public static class KernelPthreadCompatExports
         public SemaphoreSlim Semaphore { get; } = new(1, 1);
         public ulong OwnerThreadId { get; set; }
         public int RecursionCount { get; set; }
-        public int Type { get; set; } = MutexTypeDefault;
+        public int Type { get; set; } = MutexTypeErrorCheck;
         public int Protocol { get; set; }
     }
 
@@ -541,7 +540,7 @@ public static class KernelPthreadCompatExports
                     return (int)OrbisGen2Result.ORBIS_GEN2_OK;
                 }
 
-                if (state.Type is MutexTypeDefault or MutexTypeNormal or MutexTypeAdaptiveNp)
+                if (state.Type is MutexTypeNormal or MutexTypeAdaptiveNp)
                 {
                     if (tryOnly)
                     {
@@ -676,7 +675,7 @@ public static class KernelPthreadCompatExports
             return (int)OrbisGen2Result.ORBIS_GEN2_ERROR_MEMORY_FAULT;
         }
 
-        var initialState = new PthreadMutexAttrState(MutexTypeDefault, 0);
+        var initialState = new PthreadMutexAttrState(MutexTypeErrorCheck, 0);
         if (!WriteMutexAttrObject(ctx, handle, initialState))
         {
             return (int)OrbisGen2Result.ORBIS_GEN2_ERROR_MEMORY_FAULT;
@@ -735,7 +734,7 @@ public static class KernelPthreadCompatExports
         {
             if (!_mutexAttrStates.TryGetValue(resolvedAddress, out var state))
             {
-                state = new PthreadMutexAttrState(MutexTypeDefault, 0);
+                state = new PthreadMutexAttrState(MutexTypeErrorCheck, 0);
             }
 
             updatedState = state with { Type = NormalizeMutexType(type) };
@@ -764,7 +763,7 @@ public static class KernelPthreadCompatExports
         {
             if (!_mutexAttrStates.TryGetValue(resolvedAddress, out var state))
             {
-                state = new PthreadMutexAttrState(MutexTypeDefault, 0);
+                state = new PthreadMutexAttrState(MutexTypeErrorCheck, 0);
             }
 
             updatedState = state with { Protocol = protocol };
@@ -847,7 +846,7 @@ public static class KernelPthreadCompatExports
             return false;
         }
 
-        return CreateImplicitMutexState(ctx, mutexAddress, MutexTypeDefault, out resolvedAddress, out state);
+        return CreateImplicitMutexState(ctx, mutexAddress, MutexTypeErrorCheck, out resolvedAddress, out state);
     }
 
     private static ulong ResolveMutexAttrHandle(CpuContext ctx, ulong attrAddress)
@@ -891,7 +890,7 @@ public static class KernelPthreadCompatExports
         {
             return _mutexAttrStates.TryGetValue(resolvedAddress, out var state)
                 ? state
-                : new PthreadMutexAttrState(MutexTypeDefault, 0);
+                : new PthreadMutexAttrState(MutexTypeErrorCheck, 0);
         }
     }
 
@@ -1293,12 +1292,12 @@ public static class KernelPthreadCompatExports
     {
         return type switch
         {
-            0 => MutexTypeDefault,
+            0 => MutexTypeErrorCheck,
             1 => MutexTypeErrorCheck,
             2 => MutexTypeRecursive,
             3 => MutexTypeNormal,
             4 => MutexTypeAdaptiveNp,
-            _ => MutexTypeDefault,
+            _ => MutexTypeErrorCheck,
         };
     }
 

src/SharpEmu.Libs/Kernel/KernelPthreadExtendedCompatExports.cs

@@ -13,13 +13,13 @@ namespace SharpEmu.Libs.Kernel;
 public static class KernelPthreadExtendedCompatExports
 {
     private const int DefaultThreadPriority = 700;
-    private const ulong DefaultThreadAffinityMask = ulong.MaxValue;
+    private const ulong DefaultThreadAffinityMask = 0x7FUL;
     private const int DefaultDetachState = 0;
     private const ulong DefaultGuardSize = 0x1000UL;
     private const ulong DefaultStackSize = 0x1_00000UL;
-    private const int DefaultInheritSched = 0;
-    private const int DefaultSchedPolicy = 0;
-    private const int DefaultSchedPriority = 0;
+    private const int DefaultInheritSched = 4;
+    private const int DefaultSchedPolicy = 1;
+    private const int DefaultSchedPriority = DefaultThreadPriority;
     private const ulong SyntheticRwlockHandleBase = 0x00006003_0000_0000;
     private const ulong SyntheticPthreadAttrHandleBase = 0x00006004_0000_0000;
 
@@ -34,6 +34,48 @@ public static class KernelPthreadExtendedCompatExports
 
     private static readonly ConcurrentDictionary<ulong, ConcurrentDictionary<int, ulong>> _threadLocalSpecific = new();
 
+    internal static void GetThreadStartScheduling(
+        CpuContext ctx,
+        ulong attrAddress,
+        out int priority,
+        out ulong affinityMask)
+    {
+        if (attrAddress == 0)
+        {
+            priority = DefaultThreadPriority;
+            affinityMask = DefaultThreadAffinityMask;
+            return;
+        }
+
+        var resolvedAddress = ResolvePthreadAttrHandle(ctx, attrAddress);
+        lock (_stateGate)
+        {
+            var attributes = GetOrCreateAttrStateLocked(resolvedAddress);
+            priority = attributes.SchedPriority;
+            affinityMask = attributes.AffinityMask;
+        }
+    }
+
+    internal static void RegisterThreadStart(
+        ulong thread,
+        string name,
+        int priority,
+        ulong affinityMask)
+    {
+        lock (_stateGate)
+        {
+            var state = GetOrCreateThreadStateLocked(thread);
+            state.Name = name;
+            state.Priority = priority;
+            state.AffinityMask = affinityMask;
+            state.Attributes = state.Attributes with
+            {
+                SchedPriority = priority,
+                AffinityMask = affinityMask,
+            };
+        }
+    }
+
     private sealed class ThreadState
     {
         public string Name { get; set; } = string.Empty;

src/SharpEmu.Libs/Np/NpEntitlementAccessExports.cs

@@ -0,0 +1,76 @@
+// Copyright (C) 2026 SharpEmu Emulator Project
+// SPDX-License-Identifier: GPL-2.0-or-later
+
+using SharpEmu.HLE;
+
+namespace SharpEmu.Libs.Np;
+
+public static class NpEntitlementAccessExports
+{
+    private const int BootParamClearSize = 0x20;
+    private const int EmptyAddcontInfoListSize = 0x10;
+
+    [SysAbiExport(
+        Nid = "jO8DM8oyego",
+        ExportName = "sceNpEntitlementAccessInitialize",
+        Target = Generation.Gen4 | Generation.Gen5,
+        LibraryName = "libSceNpEntitlementAccess")]
+    public static int NpEntitlementAccessInitialize(CpuContext ctx)
+    {
+        var initParam = ctx[CpuRegister.Rdi];
+        var bootParam = ctx[CpuRegister.Rsi];
+
+        if (bootParam != 0)
+        {
+            Span<byte> clear = stackalloc byte[BootParamClearSize];
+            clear.Clear();
+            if (!ctx.Memory.TryWrite(bootParam, clear))
+            {
+                return SetReturn(ctx, OrbisGen2Result.ORBIS_GEN2_ERROR_MEMORY_FAULT);
+            }
+        }
+
+        TraceNpEntitlementAccess($"initialize init=0x{initParam:X16} boot=0x{bootParam:X16}");
+        return SetReturn(ctx, OrbisGen2Result.ORBIS_GEN2_OK);
+    }
+
+    [SysAbiExport(
+        Nid = "TFyU+KFBv54",
+        ExportName = "sceNpEntitlementAccessGetAddcontEntitlementInfoList",
+        Target = Generation.Gen4 | Generation.Gen5,
+        LibraryName = "libSceNpEntitlementAccess")]
+    public static int NpEntitlementAccessGetAddcontEntitlementInfoList(CpuContext ctx)
+    {
+        var listAddress = ctx[CpuRegister.Rsi];
+        if (listAddress != 0)
+        {
+            Span<byte> emptyList = stackalloc byte[EmptyAddcontInfoListSize];
+            emptyList.Clear();
+            if (!ctx.Memory.TryWrite(listAddress, emptyList))
+            {
+                return SetReturn(ctx, OrbisGen2Result.ORBIS_GEN2_ERROR_MEMORY_FAULT);
+            }
+        }
+
+        TraceNpEntitlementAccess(
+            $"get_addcont_info_list service=0x{ctx[CpuRegister.Rdi]:X16} list=0x{listAddress:X16} " +
+            $"max={ctx[CpuRegister.Rdx]} flags=0x{ctx[CpuRegister.Rcx]:X16} -> empty");
+        return SetReturn(ctx, OrbisGen2Result.ORBIS_GEN2_OK);
+    }
+
+    private static int SetReturn(CpuContext ctx, OrbisGen2Result result)
+    {
+        ctx[CpuRegister.Rax] = unchecked((ulong)(int)result);
+        return (int)result;
+    }
+
+    private static void TraceNpEntitlementAccess(string message)
+    {
+        if (!string.Equals(Environment.GetEnvironmentVariable("SHARPEMU_LOG_NP"), "1", StringComparison.Ordinal))
+        {
+            return;
+        }
+
+        Console.Error.WriteLine($"[LOADER][TRACE] np.entitlement.{message}");
+    }
+}

src/SharpEmu.Libs/PlayGo/PlayGoExports.cs

@@ -29,6 +29,7 @@ public static class PlayGoExports
     private const int PlayGoInstallSpeedTrickle = 1;
     private const int PlayGoInstallSpeedFull = 2;
     private const uint MaxPlayGoQueryEntries = 0x4000;
+    private const uint PlayGoAllEntriesSentinel = uint.MaxValue;
 
     private static readonly Regex ChunkIdPattern = new(
         @"<chunk\s+[^>]*\bid\s*=\s*""(?<id>\d+)""",
@@ -45,6 +46,7 @@ public static class PlayGoExports
     private static int _installSpeed = PlayGoInstallSpeedTrickle;
     private static ulong _languageMask = ulong.MaxValue;
     private static int _unknownChunkDiagnostics;
+    private static int _locusTraceDiagnostics;
 
     [SysAbiExport(
         Nid = "ts6GlZOKRrE",
@@ -94,6 +96,7 @@ public static class PlayGoExports
             _languageMask = ulong.MaxValue;
             _opened = false;
             _initialized = true;
+            TracePlayGo($"initialize chunks={_metadata.ChunkIds.Length} available={_metadata.Available}");
         }
 
         return (int)OrbisGen2Result.ORBIS_GEN2_OK;
@@ -131,6 +134,7 @@ public static class PlayGoExports
             }
 
             _opened = true;
+            TracePlayGo($"open handle={PlayGoHandle} chunks={_metadata.ChunkIds.Length}");
         }
 
         Span<byte> handleBytes = stackalloc byte[sizeof(uint)];
@@ -229,6 +233,7 @@ public static class PlayGoExports
         var availableEntries = chunkIds.Length == 0 ? 1u : (uint)chunkIds.Length;
         if (outChunkIdList == 0)
         {
+            TracePlayGo($"get_chunk_id count_only entries={availableEntries} out_entries=0x{outEntries:X16}");
             return TryWriteUInt32(ctx, outEntries, availableEntries)
                 ? (int)OrbisGen2Result.ORBIS_GEN2_OK
                 : (int)OrbisGen2Result.ORBIS_GEN2_ERROR_MEMORY_FAULT;
@@ -237,6 +242,7 @@ public static class PlayGoExports
         var entriesToWrite = Math.Min(numberOfEntries, availableEntries);
         if (entriesToWrite > MaxPlayGoQueryEntries)
         {
+            TracePlayGo($"get_chunk_id bad_size requested={numberOfEntries} available={availableEntries}");
             return OrbisPlayGoErrorBadSize;
         }
 
@@ -249,6 +255,7 @@ public static class PlayGoExports
             }
         }
 
+        TracePlayGo($"get_chunk_id write requested={numberOfEntries} wrote={entriesToWrite} available={availableEntries}");
         return TryWriteUInt32(ctx, outEntries, entriesToWrite)
             ? (int)OrbisGen2Result.ORBIS_GEN2_OK
             : (int)OrbisGen2Result.ORBIS_GEN2_ERROR_MEMORY_FAULT;
@@ -279,6 +286,7 @@ public static class PlayGoExports
 
         if (numberOfEntries == 0 || numberOfEntries > MaxPlayGoQueryEntries)
         {
+            TracePlayGo($"get_eta bad_size entries={numberOfEntries} chunk_ids=0x{chunkIds:X16} out=0x{outEta:X16}");
             return OrbisPlayGoErrorBadSize;
         }
 
@@ -376,8 +384,15 @@ public static class PlayGoExports
             return OrbisPlayGoErrorBadPointer;
         }
 
+        if (numberOfEntries == PlayGoAllEntriesSentinel)
+        {
+            TracePlayGo($"get_locus sentinel entries={numberOfEntries} chunk_ids=0x{chunkIds:X16} out=0x{outLoci:X16}");
+            return (int)OrbisGen2Result.ORBIS_GEN2_OK;
+        }
+
         if (numberOfEntries == 0 || numberOfEntries > MaxPlayGoQueryEntries)
         {
+            TracePlayGo($"get_locus bad_size entries={numberOfEntries} chunk_ids=0x{chunkIds:X16} out=0x{outLoci:X16}");
             return OrbisPlayGoErrorBadSize;
         }
 
@@ -408,6 +423,7 @@ public static class PlayGoExports
             loci[i] = PlayGoLocusLocalFast;
         }
 
+        TracePlayGoLocus(numberOfEntries, chunkIds, outLoci);
         return ctx.Memory.TryWrite(outLoci, loci)
             ? (int)OrbisGen2Result.ORBIS_GEN2_OK
             : (int)OrbisGen2Result.ORBIS_GEN2_ERROR_MEMORY_FAULT;
@@ -438,6 +454,7 @@ public static class PlayGoExports
 
         if (numberOfEntries == 0 || numberOfEntries > MaxPlayGoQueryEntries)
         {
+            TracePlayGo($"get_progress bad_size entries={numberOfEntries} chunk_ids=0x{chunkIds:X16} out=0x{outProgress:X16}");
             return OrbisPlayGoErrorBadSize;
         }
 
@@ -447,6 +464,7 @@ public static class PlayGoExports
             return chunkError;
         }
 
+        TracePlayGo($"get_progress entries={numberOfEntries}");
         Span<byte> progress = stackalloc byte[sizeof(ulong) * 2];
         BinaryPrimitives.WriteUInt64LittleEndian(progress, 0);
         BinaryPrimitives.WriteUInt64LittleEndian(progress[sizeof(ulong)..], 0);
@@ -480,9 +498,11 @@ public static class PlayGoExports
 
         if (numberOfEntries == 0)
         {
+            TracePlayGo("get_todo bad_size entries=0");
             return OrbisPlayGoErrorBadSize;
         }
 
+        TracePlayGo($"get_todo requested={numberOfEntries} wrote=0");
         return TryWriteUInt32(ctx, outEntries, 0)
             ? (int)OrbisGen2Result.ORBIS_GEN2_OK
             : (int)OrbisGen2Result.ORBIS_GEN2_ERROR_MEMORY_FAULT;
@@ -749,6 +769,29 @@ public static class PlayGoExports
         return ctx.Memory.TryWrite(address, buffer);
     }
 
+    private static void TracePlayGo(string message)
+    {
+        if (string.Equals(Environment.GetEnvironmentVariable("SHARPEMU_LOG_PLAYGO"), "1", StringComparison.Ordinal))
+        {
+            Console.Error.WriteLine($"[LOADER][TRACE] playgo.{message}");
+        }
+    }
+
+    private static void TracePlayGoLocus(uint entries, ulong chunkIds, ulong outLoci)
+    {
+        if (!string.Equals(Environment.GetEnvironmentVariable("SHARPEMU_LOG_PLAYGO"), "1", StringComparison.Ordinal))
+        {
+            return;
+        }
+
+        var count = Interlocked.Increment(ref _locusTraceDiagnostics);
+        if (entries != 1 || count <= 32 || count % 1000 == 0)
+        {
+            Console.Error.WriteLine(
+                $"[LOADER][TRACE] playgo.get_locus entries={entries} chunk_ids=0x{chunkIds:X16} out=0x{outLoci:X16}");
+        }
+    }
+
     private sealed record PlayGoMetadata(bool Available, ushort[] ChunkIds)
     {
         public static readonly PlayGoMetadata Empty = new(false, Array.Empty<ushort>());

src/SharpEmu.Libs/SaveData/SaveDataDialogExports.cs

@@ -0,0 +1,213 @@
+// Copyright (C) 2026 SharpEmu Emulator Project
+// SPDX-License-Identifier: GPL-2.0-or-later
+
+using SharpEmu.HLE;
+using System.Buffers.Binary;
+using System.Threading;
+
+namespace SharpEmu.Libs.SaveData;
+
+public static class SaveDataDialogExports
+{
+    private const int StatusNone = 0;
+    private const int StatusInitialized = 1;
+    private const int StatusRunning = 2;
+    private const int StatusFinished = 3;
+
+    private const int ErrorOk = 0;
+    private const int ErrorNotInitialized = unchecked((int)0x80B80003);
+    private const int ErrorAlreadyInitialized = unchecked((int)0x80B80004);
+    private const int ErrorNotFinished = unchecked((int)0x80B80005);
+    private const int ErrorNotRunning = unchecked((int)0x80B8000B);
+    private const int ErrorArgNull = unchecked((int)0x80B8000D);
+
+    private const int ResultSize = 0x48;
+    private static int _status;
+    private static int _lastMode;
+    private static ulong _lastUserData;
+
+    [SysAbiExport(
+        Nid = "s9e3+YpRnzw",
+        ExportName = "sceSaveDataDialogInitialize",
+        Target = Generation.Gen4 | Generation.Gen5,
+        LibraryName = "libSceSaveDataDialog")]
+    public static int SaveDataDialogInitialize(CpuContext ctx)
+    {
+        if (Interlocked.CompareExchange(ref _status, StatusInitialized, StatusNone) != StatusNone)
+        {
+            return SetReturn(ctx, ErrorAlreadyInitialized);
+        }
+
+        return SetReturn(ctx, ErrorOk);
+    }
+
+    [SysAbiExport(
+        Nid = "4tPhsP6FpDI",
+        ExportName = "sceSaveDataDialogOpen",
+        Target = Generation.Gen4 | Generation.Gen5,
+        LibraryName = "libSceSaveDataDialog")]
+    public static int SaveDataDialogOpen(CpuContext ctx)
+    {
+        var paramAddress = ctx[CpuRegister.Rdi];
+        if (paramAddress == 0)
+        {
+            return SetReturn(ctx, ErrorArgNull);
+        }
+
+        if (_status is not (StatusInitialized or StatusFinished))
+        {
+            return SetReturn(ctx, ErrorNotInitialized);
+        }
+
+        _lastMode = TryReadInt32(ctx, paramAddress, out var mode) ? mode : 0;
+        _lastUserData = TryReadUInt64(ctx, paramAddress + 0xC8, out var userData) ? userData : 0;
+
+        // There is no host save dialog yet. Complete immediately with OK so
+        // guest polling sees a finished dialog instead of spinning forever.
+        Interlocked.Exchange(ref _status, StatusFinished);
+        TraceSaveDataDialog($"open mode={_lastMode} userData=0x{_lastUserData:X16} -> finished");
+        return SetReturn(ctx, ErrorOk);
+    }
+
+    [SysAbiExport(
+        Nid = "ERKzksauAJA",
+        ExportName = "sceSaveDataDialogGetStatus",
+        Target = Generation.Gen4 | Generation.Gen5,
+        LibraryName = "libSceSaveDataDialog")]
+    public static int SaveDataDialogGetStatus(CpuContext ctx) => SetReturn(ctx, Volatile.Read(ref _status));
+
+    [SysAbiExport(
+        Nid = "KK3Bdg1RWK0",
+        ExportName = "sceSaveDataDialogUpdateStatus",
+        Target = Generation.Gen4 | Generation.Gen5,
+        LibraryName = "libSceSaveDataDialog")]
+    public static int SaveDataDialogUpdateStatus(CpuContext ctx) => SetReturn(ctx, Volatile.Read(ref _status));
+
+    [SysAbiExport(
+        Nid = "en7gNVnh878",
+        ExportName = "sceSaveDataDialogIsReadyToDisplay",
+        Target = Generation.Gen4 | Generation.Gen5,
+        LibraryName = "libSceSaveDataDialog")]
+    public static int SaveDataDialogIsReadyToDisplay(CpuContext ctx) => SetReturn(ctx, 1);
+
+    [SysAbiExport(
+        Nid = "yEiJ-qqr6Cg",
+        ExportName = "sceSaveDataDialogGetResult",
+        Target = Generation.Gen4 | Generation.Gen5,
+        LibraryName = "libSceSaveDataDialog")]
+    public static int SaveDataDialogGetResult(CpuContext ctx)
+    {
+        var resultAddress = ctx[CpuRegister.Rdi];
+        if (resultAddress == 0)
+        {
+            return SetReturn(ctx, ErrorArgNull);
+        }
+
+        if (Volatile.Read(ref _status) != StatusFinished)
+        {
+            return SetReturn(ctx, ErrorNotFinished);
+        }
+
+        Span<byte> result = stackalloc byte[ResultSize];
+        result.Clear();
+        BinaryPrimitives.WriteInt32LittleEndian(result[0x00..], _lastMode);
+        BinaryPrimitives.WriteInt32LittleEndian(result[0x04..], 0);
+        BinaryPrimitives.WriteInt32LittleEndian(result[0x08..], 0);
+        BinaryPrimitives.WriteUInt64LittleEndian(result[0x20..], _lastUserData);
+
+        if (!ctx.Memory.TryWrite(resultAddress, result))
+        {
+            return SetReturn(ctx, (int)OrbisGen2Result.ORBIS_GEN2_ERROR_MEMORY_FAULT);
+        }
+
+        return SetReturn(ctx, ErrorOk);
+    }
+
+    [SysAbiExport(
+        Nid = "fH46Lag88XY",
+        ExportName = "sceSaveDataDialogClose",
+        Target = Generation.Gen4 | Generation.Gen5,
+        LibraryName = "libSceSaveDataDialog")]
+    public static int SaveDataDialogClose(CpuContext ctx)
+    {
+        if (Interlocked.CompareExchange(ref _status, StatusFinished, StatusRunning) != StatusRunning)
+        {
+            return SetReturn(ctx, ErrorNotRunning);
+        }
+
+        return SetReturn(ctx, ErrorOk);
+    }
+
+    [SysAbiExport(
+        Nid = "YuH2FA7azqQ",
+        ExportName = "sceSaveDataDialogTerminate",
+        Target = Generation.Gen4 | Generation.Gen5,
+        LibraryName = "libSceSaveDataDialog")]
+    public static int SaveDataDialogTerminate(CpuContext ctx)
+    {
+        if (Interlocked.Exchange(ref _status, StatusNone) == StatusNone)
+        {
+            return SetReturn(ctx, ErrorNotInitialized);
+        }
+
+        _lastMode = 0;
+        _lastUserData = 0;
+        return SetReturn(ctx, ErrorOk);
+    }
+
+    [SysAbiExport(
+        Nid = "V-uEeFKARJU",
+        ExportName = "sceSaveDataDialogProgressBarInc",
+        Target = Generation.Gen4 | Generation.Gen5,
+        LibraryName = "libSceSaveDataDialog")]
+    public static int SaveDataDialogProgressBarInc(CpuContext ctx) => SetReturn(ctx, ErrorOk);
+
+    [SysAbiExport(
+        Nid = "hay1CfTmLyA",
+        ExportName = "sceSaveDataDialogProgressBarSetValue",
+        Target = Generation.Gen4 | Generation.Gen5,
+        LibraryName = "libSceSaveDataDialog")]
+    public static int SaveDataDialogProgressBarSetValue(CpuContext ctx) => SetReturn(ctx, ErrorOk);
+
+    private static bool TryReadInt32(CpuContext ctx, ulong address, out int value)
+    {
+        value = 0;
+        Span<byte> bytes = stackalloc byte[sizeof(int)];
+        if (!ctx.Memory.TryRead(address, bytes))
+        {
+            return false;
+        }
+
+        value = BinaryPrimitives.ReadInt32LittleEndian(bytes);
+        return true;
+    }
+
+    private static bool TryReadUInt64(CpuContext ctx, ulong address, out ulong value)
+    {
+        value = 0;
+        Span<byte> bytes = stackalloc byte[sizeof(ulong)];
+        if (!ctx.Memory.TryRead(address, bytes))
+        {
+            return false;
+        }
+
+        value = BinaryPrimitives.ReadUInt64LittleEndian(bytes);
+        return true;
+    }
+
+    private static int SetReturn(CpuContext ctx, int result)
+    {
+        ctx[CpuRegister.Rax] = unchecked((ulong)result);
+        return result;
+    }
+
+    private static void TraceSaveDataDialog(string message)
+    {
+        if (!string.Equals(Environment.GetEnvironmentVariable("SHARPEMU_LOG_SAVEDATA"), "1", StringComparison.Ordinal))
+        {
+            return;
+        }
+
+        Console.Error.WriteLine($"[LOADER][TRACE] save_data_dialog.{message}");
+    }
+}

src/SharpEmu.Libs/SaveData/SaveDataExports.cs

@@ -0,0 +1,539 @@
+// Copyright (C) 2026 SharpEmu Emulator Project
+// SPDX-License-Identifier: GPL-2.0-or-later
+
+using SharpEmu.HLE;
+using SharpEmu.Libs.Kernel;
+using System.Buffers.Binary;
+using System.Text;
+
+namespace SharpEmu.Libs.SaveData;
+
+public static class SaveDataExports
+{
+    private const int OrbisSaveDataErrorParameter = unchecked((int)0x809F0000);
+    private const int OrbisSaveDataErrorExists = unchecked((int)0x809F0007);
+    private const int OrbisSaveDataErrorNotFound = unchecked((int)0x809F0008);
+    private const int OrbisSaveDataErrorInternal = unchecked((int)0x809F000B);
+    private const int SaveDataTitleIdSize = 10;
+    private const int SaveDataDirNameSize = 32;
+    private const int SaveDataParamSize = 0x530;
+    private const int SaveDataSearchInfoSize = 0x30;
+    private const ulong ResultHitNumOffset = 0x00;
+    private const ulong ResultDirNamesOffset = 0x08;
+    private const ulong ResultDirNamesNumOffset = 0x10;
+    private const ulong ResultSetNumOffset = 0x14;
+    private const ulong ResultParamsOffset = 0x18;
+    private const ulong ResultInfosOffset = 0x20;
+    private const uint SortKeyFreeBlocks = 5;
+    private const uint SortOrderDescent = 1;
+    private const uint MountModeCreate = 1u << 2;
+    private const uint MountModeCreate2 = 1u << 5;
+    private const int MountResultSize = 0x40;
+    private static readonly object _stateGate = new();
+    private static string? _titleId;
+
+    public static void ConfigureApplicationInfo(string? titleId)
+    {
+        lock (_stateGate)
+        {
+            _titleId = string.IsNullOrWhiteSpace(titleId) ? null : SanitizePathSegment(titleId.Trim());
+        }
+    }
+
+    [SysAbiExport(
+        Nid = "TywrFKCoLGY",
+        ExportName = "sceSaveDataInitialize3",
+        Target = Generation.Gen4 | Generation.Gen5,
+        LibraryName = "libSceSaveData")]
+    public static int SaveDataInitialize3(CpuContext ctx)
+    {
+        try
+        {
+            Directory.CreateDirectory(ResolveSaveDataRoot());
+            return SetReturn(ctx, 0);
+        }
+        catch (IOException)
+        {
+            return SetReturn(ctx, OrbisSaveDataErrorInternal);
+        }
+        catch (UnauthorizedAccessException)
+        {
+            return SetReturn(ctx, OrbisSaveDataErrorInternal);
+        }
+    }
+
+    [SysAbiExport(
+        Nid = "dyIhnXq-0SM",
+        ExportName = "sceSaveDataDirNameSearch",
+        Target = Generation.Gen4 | Generation.Gen5,
+        LibraryName = "libSceSaveData")]
+    public static int SaveDataDirNameSearch(CpuContext ctx)
+    {
+        var condAddress = ctx[CpuRegister.Rdi];
+        var resultAddress = ctx[CpuRegister.Rsi];
+        if (condAddress == 0 || resultAddress == 0)
+        {
+            return SetReturn(ctx, OrbisSaveDataErrorParameter);
+        }
+
+        if (!TryReadSearchCond(ctx, condAddress, out var cond) ||
+            !TryReadSearchResult(ctx, resultAddress, out var result))
+        {
+            return SetReturn(ctx, (int)OrbisGen2Result.ORBIS_GEN2_ERROR_MEMORY_FAULT);
+        }
+
+        if (cond.UserId < 0 || cond.SortKey > SortKeyFreeBlocks || cond.SortOrder > SortOrderDescent)
+        {
+            return SetReturn(ctx, OrbisSaveDataErrorParameter);
+        }
+
+        try
+        {
+            string titleId;
+            if (cond.TitleIdAddress == 0)
+            {
+                titleId = ResolveConfiguredTitleId();
+            }
+            else if (!TryReadFixedAscii(ctx, cond.TitleIdAddress, SaveDataTitleIdSize, out titleId))
+            {
+                return SetReturn(ctx, (int)OrbisGen2Result.ORBIS_GEN2_ERROR_MEMORY_FAULT);
+            }
+
+            var root = ResolveTitleSaveRoot(cond.UserId, titleId);
+            var entries = Directory.Exists(root)
+                ? EnumerateSaveDirectories(root, cond.Pattern)
+                : [];
+
+            entries = SortEntries(entries, cond.SortKey, cond.SortOrder);
+            var setNum = result.DirNamesNum == 0
+                ? 0
+                : Math.Min(result.DirNamesNum, entries.Count);
+            if (!TryWriteUInt32(ctx, resultAddress + ResultHitNumOffset, checked((uint)entries.Count)) ||
+                !TryWriteUInt32(ctx, resultAddress + ResultSetNumOffset, checked((uint)setNum)))
+            {
+                return SetReturn(ctx, (int)OrbisGen2Result.ORBIS_GEN2_ERROR_MEMORY_FAULT);
+            }
+
+            if (setNum == 0)
+            {
+                TraceSaveData($"dir_name_search user={cond.UserId} title={titleId} hits={entries.Count} set=0 root='{root}'");
+                return SetReturn(ctx, 0);
+            }
+
+            if (result.DirNamesAddress == 0)
+            {
+                return SetReturn(ctx, OrbisSaveDataErrorParameter);
+            }
+
+            for (var i = 0; i < setNum; i++)
+            {
+                var entry = entries[i];
+                if (!TryWriteFixedAscii(
+                        ctx,
+                        result.DirNamesAddress + ((ulong)i * SaveDataDirNameSize),
+                        SaveDataDirNameSize,
+                        entry.Name) ||
+                    (result.ParamsAddress != 0 &&
+                     !TryWriteParam(ctx, result.ParamsAddress + ((ulong)i * SaveDataParamSize), entry)) ||
+                    (result.InfosAddress != 0 &&
+                     !TryWriteSearchInfo(ctx, result.InfosAddress + ((ulong)i * SaveDataSearchInfoSize), entry)))
+                {
+                    return SetReturn(ctx, (int)OrbisGen2Result.ORBIS_GEN2_ERROR_MEMORY_FAULT);
+                }
+            }
+
+            TraceSaveData($"dir_name_search user={cond.UserId} title={titleId} hits={entries.Count} set={setNum} root='{root}'");
+            return SetReturn(ctx, 0);
+        }
+        catch (IOException)
+        {
+            return SetReturn(ctx, OrbisSaveDataErrorInternal);
+        }
+        catch (UnauthorizedAccessException)
+        {
+            return SetReturn(ctx, OrbisSaveDataErrorInternal);
+        }
+    }
+
+    [SysAbiExport(
+        Nid = "ZP4e7rlzOUk",
+        ExportName = "sceSaveDataMount3",
+        Target = Generation.Gen5,
+        LibraryName = "libSceSaveData")]
+    public static int SaveDataMount3(CpuContext ctx)
+    {
+        var mountAddress = ctx[CpuRegister.Rdi];
+        var resultAddress = ctx[CpuRegister.Rsi];
+        if (mountAddress == 0 || resultAddress == 0)
+        {
+            return SetReturn(ctx, OrbisSaveDataErrorParameter);
+        }
+
+        if (!TryReadInt32(ctx, mountAddress, out var userId) ||
+            !ctx.TryReadUInt64(mountAddress + 0x08, out var dirNameAddress) ||
+            !ctx.TryReadUInt64(mountAddress + 0x10, out var blocks) ||
+            !ctx.TryReadUInt64(mountAddress + 0x18, out var systemBlocks) ||
+            !TryReadUInt32(ctx, mountAddress + 0x20, out var mountMode) ||
+            !TryReadUInt32(ctx, mountAddress + 0x24, out var resource) ||
+            !TryReadUInt32(ctx, mountAddress + 0x28, out var mode) ||
+            dirNameAddress == 0 ||
+            !TryReadFixedAscii(ctx, dirNameAddress, SaveDataDirNameSize, out var dirName))
+        {
+            return SetReturn(ctx, (int)OrbisGen2Result.ORBIS_GEN2_ERROR_MEMORY_FAULT);
+        }
+
+        if (userId < 0 || string.IsNullOrWhiteSpace(dirName))
+        {
+            return SetReturn(ctx, OrbisSaveDataErrorParameter);
+        }
+
+        try
+        {
+            var titleId = ResolveConfiguredTitleId();
+            var savePath = Path.Combine(
+                ResolveTitleSaveRoot(userId, titleId),
+                SanitizePathSegment(dirName));
+            var existed = Directory.Exists(savePath);
+            var create = (mountMode & MountModeCreate) != 0;
+            var createIfMissing = (mountMode & MountModeCreate2) != 0;
+
+            if (!existed && !create && !createIfMissing)
+            {
+                return SetReturn(ctx, OrbisSaveDataErrorNotFound);
+            }
+
+            if (existed && create)
+            {
+                return SetReturn(ctx, OrbisSaveDataErrorExists);
+            }
+
+            if (!existed)
+            {
+                Directory.CreateDirectory(savePath);
+            }
+
+            const string mountPoint = "/savedata0";
+            KernelMemoryCompatExports.RegisterGuestPathMount(mountPoint, savePath);
+
+            Span<byte> result = stackalloc byte[MountResultSize];
+            result.Clear();
+            WriteAscii(result[..16], mountPoint);
+            BinaryPrimitives.WriteUInt32LittleEndian(result[0x1C..], createIfMissing && !existed ? 1u : 0u);
+            if (!ctx.Memory.TryWrite(resultAddress, result))
+            {
+                return SetReturn(ctx, (int)OrbisGen2Result.ORBIS_GEN2_ERROR_MEMORY_FAULT);
+            }
+
+            TraceSaveData(
+                $"mount3 user={userId} title={titleId} dir={dirName} blocks={blocks} " +
+                $"system_blocks={systemBlocks} mount_mode=0x{mountMode:X} resource={resource} mode={mode} " +
+                $"mount_point={mountPoint} created={!existed} root='{savePath}'");
+            return SetReturn(ctx, 0);
+        }
+        catch (IOException)
+        {
+            return SetReturn(ctx, OrbisSaveDataErrorInternal);
+        }
+        catch (UnauthorizedAccessException)
+        {
+            return SetReturn(ctx, OrbisSaveDataErrorInternal);
+        }
+        catch (ArgumentException)
+        {
+            return SetReturn(ctx, OrbisSaveDataErrorParameter);
+        }
+    }
+
+    private static bool TryReadSearchCond(CpuContext ctx, ulong address, out SearchCond cond)
+    {
+        cond = default;
+        if (!TryReadInt32(ctx, address, out var userId) ||
+            !ctx.TryReadUInt64(address + 0x08, out var titleIdAddress) ||
+            !ctx.TryReadUInt64(address + 0x10, out var dirNameAddress) ||
+            !TryReadUInt32(ctx, address + 0x18, out var sortKey) ||
+            !TryReadUInt32(ctx, address + 0x1C, out var sortOrder))
+        {
+            return false;
+        }
+
+        string pattern;
+        if (dirNameAddress == 0)
+        {
+            pattern = string.Empty;
+        }
+        else if (!TryReadFixedAscii(ctx, dirNameAddress, SaveDataDirNameSize, out pattern))
+        {
+            return false;
+        }
+
+        cond = new SearchCond(userId, titleIdAddress, pattern, sortKey, sortOrder);
+        return true;
+    }
+
+    private static bool TryReadSearchResult(CpuContext ctx, ulong address, out SearchResult result)
+    {
+        result = default;
+        if (!ctx.TryReadUInt64(address + ResultDirNamesOffset, out var dirNamesAddress) ||
+            !TryReadUInt32(ctx, address + ResultDirNamesNumOffset, out var dirNamesNum) ||
+            !ctx.TryReadUInt64(address + ResultParamsOffset, out var paramsAddress) ||
+            !ctx.TryReadUInt64(address + ResultInfosOffset, out var infosAddress))
+        {
+            return false;
+        }
+
+        result = new SearchResult(dirNamesAddress, dirNamesNum, paramsAddress, infosAddress);
+        return true;
+    }
+
+    private static List<SaveEntry> EnumerateSaveDirectories(string root, string pattern)
+    {
+        var entries = new List<SaveEntry>();
+        foreach (var directory in Directory.EnumerateDirectories(root))
+        {
+            var name = Path.GetFileName(directory);
+            if (string.IsNullOrWhiteSpace(name) ||
+                name.StartsWith("sce_", StringComparison.OrdinalIgnoreCase) ||
+                (!string.IsNullOrEmpty(pattern) && !MatchPattern(name, pattern)))
+            {
+                continue;
+            }
+
+            var info = new DirectoryInfo(directory);
+            entries.Add(new SaveEntry(name, directory, info.LastWriteTimeUtc));
+        }
+
+        return entries;
+    }
+
+    private static List<SaveEntry> SortEntries(List<SaveEntry> entries, uint sortKey, uint sortOrder)
+    {
+        IOrderedEnumerable<SaveEntry> sorted = sortKey switch
+        {
+            3 => entries.OrderBy(entry => entry.LastWriteUtc),
+            _ => entries.OrderBy(entry => entry.Name, StringComparer.OrdinalIgnoreCase),
+        };
+
+        var list = sorted.ToList();
+        if (sortOrder == SortOrderDescent)
+        {
+            list.Reverse();
+        }
+
+        return list;
+    }
+
+    private static bool TryWriteParam(CpuContext ctx, ulong address, SaveEntry entry)
+    {
+        var param = new byte[SaveDataParamSize];
+        WriteAscii(param.AsSpan(0x00, 128), "Saved Data");
+        WriteAscii(param.AsSpan(0x100, 1024), entry.Name);
+        BinaryPrimitives.WriteInt64LittleEndian(
+            param.AsSpan(0x508, sizeof(long)),
+            new DateTimeOffset(entry.LastWriteUtc).ToUnixTimeSeconds());
+        return ctx.Memory.TryWrite(address, param);
+    }
+
+    private static bool TryWriteSearchInfo(CpuContext ctx, ulong address, SaveEntry entry)
+    {
+        var size = GetDirectorySize(entry.Path);
+        var usedBlocks = checked((ulong)((size + 32767) / 32768));
+        var blocks = Math.Max(96UL, usedBlocks);
+        Span<byte> info = stackalloc byte[SaveDataSearchInfoSize];
+        info.Clear();
+        BinaryPrimitives.WriteUInt64LittleEndian(info[0x00..], blocks);
+        BinaryPrimitives.WriteUInt64LittleEndian(info[0x08..], blocks - usedBlocks);
+        return ctx.Memory.TryWrite(address, info);
+    }
+
+    private static long GetDirectorySize(string root)
+    {
+        long total = 0;
+        foreach (var file in Directory.EnumerateFiles(root, "*", SearchOption.AllDirectories))
+        {
+            total += new FileInfo(file).Length;
+        }
+
+        return total;
+    }
+
+    private static bool MatchPattern(string value, string pattern) =>
+        MatchPattern(value.AsSpan(), pattern.AsSpan());
+
+    private static bool MatchPattern(ReadOnlySpan<char> value, ReadOnlySpan<char> pattern)
+    {
+        if (pattern.IsEmpty)
+        {
+            return value.IsEmpty;
+        }
+
+        if (pattern[0] == '%')
+        {
+            for (var i = 0; i <= value.Length; i++)
+            {
+                if (MatchPattern(value[i..], pattern[1..]))
+                {
+                    return true;
+                }
+            }
+
+            return false;
+        }
+
+        if (value.IsEmpty)
+        {
+            return false;
+        }
+
+        if (pattern[0] == '_' ||
+            char.ToUpperInvariant(pattern[0]) == char.ToUpperInvariant(value[0]))
+        {
+            return MatchPattern(value[1..], pattern[1..]);
+        }
+
+        return false;
+    }
+
+    private static string ResolveTitleSaveRoot(int userId, string titleId) =>
+        Path.Combine(ResolveSaveDataRoot(), userId.ToString(), SanitizePathSegment(titleId));
+
+    private static string ResolveSaveDataRoot()
+    {
+        var configured = Environment.GetEnvironmentVariable("SHARPEMU_SAVEDATA_DIR");
+        var root = string.IsNullOrWhiteSpace(configured)
+            ? Path.Combine(Environment.CurrentDirectory, "user", "savedata")
+            : configured;
+        return Path.GetFullPath(root);
+    }
+
+    private static string ResolveConfiguredTitleId()
+    {
+        lock (_stateGate)
+        {
+            if (!string.IsNullOrWhiteSpace(_titleId))
+            {
+                return _titleId;
+            }
+        }
+
+        var app0Root = Environment.GetEnvironmentVariable("SHARPEMU_APP0_DIR");
+        var app0Name = string.IsNullOrWhiteSpace(app0Root)
+            ? null
+            : Path.GetFileName(Path.TrimEndingDirectorySeparator(app0Root));
+        if (!string.IsNullOrWhiteSpace(app0Name))
+        {
+            var candidate = app0Name.Split('-', StringSplitOptions.RemoveEmptyEntries)[0];
+            if (!string.IsNullOrWhiteSpace(candidate))
+            {
+                return SanitizePathSegment(candidate);
+            }
+        }
+
+        return "default";
+    }
+
+    private static string SanitizePathSegment(string value)
+    {
+        var invalid = Path.GetInvalidFileNameChars();
+        var sanitized = new string(value.Select(ch => invalid.Contains(ch) ? '_' : ch).ToArray());
+        return string.IsNullOrWhiteSpace(sanitized) ? "default" : sanitized;
+    }
+
+    private static bool TryReadFixedAscii(CpuContext ctx, ulong address, int length, out string value)
+    {
+        value = string.Empty;
+        Span<byte> buffer = stackalloc byte[length];
+        if (!ctx.Memory.TryRead(address, buffer))
+        {
+            return false;
+        }
+
+        var stringLength = buffer.IndexOf((byte)0);
+        if (stringLength < 0)
+        {
+            stringLength = buffer.Length;
+        }
+
+        value = Encoding.ASCII.GetString(buffer[..stringLength]);
+        return true;
+    }
+
+    private static bool TryWriteFixedAscii(CpuContext ctx, ulong address, int length, string value)
+    {
+        Span<byte> buffer = stackalloc byte[length];
+        buffer.Clear();
+        WriteAscii(buffer, value);
+        return ctx.Memory.TryWrite(address, buffer);
+    }
+
+    private static void WriteAscii(Span<byte> destination, string value)
+    {
+        var count = Math.Min(value.Length, Math.Max(0, destination.Length - 1));
+        for (var i = 0; i < count; i++)
+        {
+            var ch = value[i];
+            destination[i] = ch <= 0x7F ? (byte)ch : (byte)'?';
+        }
+    }
+
+    private static bool TryReadInt32(CpuContext ctx, ulong address, out int value)
+    {
+        Span<byte> bytes = stackalloc byte[sizeof(int)];
+        if (!ctx.Memory.TryRead(address, bytes))
+        {
+            value = 0;
+            return false;
+        }
+
+        value = BinaryPrimitives.ReadInt32LittleEndian(bytes);
+        return true;
+    }
+
+    private static bool TryReadUInt32(CpuContext ctx, ulong address, out uint value)
+    {
+        Span<byte> bytes = stackalloc byte[sizeof(uint)];
+        if (!ctx.Memory.TryRead(address, bytes))
+        {
+            value = 0;
+            return false;
+        }
+
+        value = BinaryPrimitives.ReadUInt32LittleEndian(bytes);
+        return true;
+    }
+
+    private static bool TryWriteUInt32(CpuContext ctx, ulong address, uint value)
+    {
+        Span<byte> bytes = stackalloc byte[sizeof(uint)];
+        BinaryPrimitives.WriteUInt32LittleEndian(bytes, value);
+        return ctx.Memory.TryWrite(address, bytes);
+    }
+
+    private static int SetReturn(CpuContext ctx, int result)
+    {
+        ctx[CpuRegister.Rax] = unchecked((ulong)result);
+        return result;
+    }
+
+    private static void TraceSaveData(string message)
+    {
+        if (string.Equals(Environment.GetEnvironmentVariable("SHARPEMU_LOG_SAVEDATA"), "1", StringComparison.Ordinal))
+        {
+            Console.Error.WriteLine($"[LOADER][TRACE] savedata.{message}");
+        }
+    }
+
+    private readonly record struct SearchCond(
+        int UserId,
+        ulong TitleIdAddress,
+        string Pattern,
+        uint SortKey,
+        uint SortOrder);
+
+    private readonly record struct SearchResult(
+        ulong DirNamesAddress,
+        uint DirNamesNum,
+        ulong ParamsAddress,
+        ulong InfosAddress);
+
+    private readonly record struct SaveEntry(string Name, string Path, DateTime LastWriteUtc);
+}

src/SharpEmu.Libs/SystemService/SystemServiceExports.cs

@@ -2,6 +2,7 @@
 // SPDX-License-Identifier: GPL-2.0-or-later
 
 using SharpEmu.HLE;
+using SharpEmu.Libs.VideoOut;
 using System.Buffers.Binary;
 
 namespace SharpEmu.Libs.SystemService;
@@ -85,6 +86,17 @@ public static class SystemServiceExports
             : SetReturn(ctx, (int)OrbisGen2Result.ORBIS_GEN2_ERROR_MEMORY_FAULT);
     }
 
+    [SysAbiExport(
+        Nid = "Vo5V8KAwCmk",
+        ExportName = "sceSystemServiceHideSplashScreen",
+        Target = Generation.Gen4 | Generation.Gen5,
+        LibraryName = "libSceSystemService")]
+    public static int SystemServiceHideSplashScreen(CpuContext ctx)
+    {
+        VulkanVideoPresenter.HideSplashScreen();
+        return SetReturn(ctx, 0);
+    }
+
     private static int SetReturn(CpuContext ctx, int result)
     {
         ctx[CpuRegister.Rax] = unchecked((ulong)result);

src/SharpEmu.Libs/VideoOut/VideoOutExports.cs

@@ -360,7 +360,24 @@ public static class VideoOutExports
         var bufferIndex = unchecked((int)ctx[CpuRegister.Rsi]);
         var flipMode = unchecked((int)ctx[CpuRegister.Rdx]);
         var flipArg = unchecked((long)ctx[CpuRegister.Rcx]);
-        return SubmitFlip(ctx, handle, bufferIndex, flipMode, flipArg);
+        return SubmitFlip(ctx, handle, bufferIndex, flipMode, flipArg, submitGpuImage: true);
+    }
+
+    [SysAbiExport(
+        Nid = "zgXifHT9ErY",
+        ExportName = "sceVideoOutIsFlipPending",
+        Target = Generation.Gen4 | Generation.Gen5,
+        LibraryName = "libSceVideoOut")]
+    public static int VideoOutIsFlipPending(CpuContext ctx)
+    {
+        var handle = unchecked((int)ctx[CpuRegister.Rdi]);
+        if (!TryGetPort(handle, out _))
+        {
+            return OrbisVideoOutErrorInvalidHandle;
+        }
+
+        ctx[CpuRegister.Rax] = 0;
+        return (int)OrbisGen2Result.ORBIS_GEN2_OK;
     }
 
     [SysAbiExport(
@@ -423,7 +440,7 @@ public static class VideoOutExports
     }
 
     public static int SubmitFlipFromAgc(CpuContext ctx, int handle, int bufferIndex, int flipMode, long flipArg) =>
-        SubmitFlip(ctx, handle, bufferIndex, flipMode, flipArg);
+        SubmitFlip(ctx, handle, bufferIndex, flipMode, flipArg, submitGpuImage: false);
 
     internal static void SubmitHostRgbaFrame(ReadOnlySpan<byte> rgbaFrame, uint width, uint height)
     {
@@ -731,7 +748,13 @@ public static class VideoOutExports
         return groupIndex < 0 ? groupIndex : setIndex;
     }
 
-    private static int SubmitFlip(CpuContext ctx, int handle, int bufferIndex, int flipMode, long flipArg)
+    private static int SubmitFlip(
+        CpuContext ctx,
+        int handle,
+        int bufferIndex,
+        int flipMode,
+        long flipArg,
+        bool submitGpuImage)
     {
         if (!TryGetPort(handle, out var port))
         {
@@ -759,6 +782,17 @@ public static class VideoOutExports
             flipEvents = new List<FlipEventRegistration>(port.FlipEvents);
         }
 
+        if (submitGpuImage &&
+            bufferIndex >= 0 &&
+            TryGetDisplayBufferInfo(handle, bufferIndex, out var displayBuffer))
+        {
+            _ = VulkanVideoPresenter.TrySubmitGuestImage(
+                displayBuffer.Address,
+                displayBuffer.Width,
+                displayBuffer.Height,
+                displayBuffer.PitchInPixel);
+        }
+
         if (string.Equals(
                 Environment.GetEnvironmentVariable("SHARPEMU_DUMP_VIDEOOUT"),
                 "1",