[fiber] synchronization problems have been fixed for such a titles: Demon's Souls

[ampr] new exports
[memory] trampoline fixes

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

@@ -19,7 +19,12 @@ public sealed partial class DirectExecutionBackend
 	{
 		if (!string.Equals(Environment.GetEnvironmentVariable("SHARPEMU_DISABLE_RAW_HANDLER"), "1", StringComparison.Ordinal))
 		{
-			_rawExceptionHandler = (nint)AddVectoredExceptionHandler(1u, RawVectoredHandlerPtrManaged);
+			_rawExceptionHandlerStub = CreateExceptionHandlerTrampoline(RawVectoredHandlerPtrManaged);
+			if (_rawExceptionHandlerStub == 0)
+			{
+				throw new InvalidOperationException("Failed to create raw exception handler trampoline");
+			}
+			_rawExceptionHandler = (nint)AddVectoredExceptionHandler(1u, _rawExceptionHandlerStub);
 			Console.Error.WriteLine($"[LOADER][INFO] Raw exception handler installed: 0x{_rawExceptionHandler:X16}");
 		}
 		else
@@ -29,12 +34,22 @@ public sealed partial class DirectExecutionBackend
 
 		_handlerDelegate = VectoredHandler;
 		_handlerHandle = GCHandle.Alloc(_handlerDelegate);
-		_exceptionHandler = (nint)AddVectoredExceptionHandler(1u, Marshal.GetFunctionPointerForDelegate(_handlerDelegate));
+		_exceptionHandlerStub = CreateExceptionHandlerTrampoline(Marshal.GetFunctionPointerForDelegate(_handlerDelegate));
+		if (_exceptionHandlerStub == 0)
+		{
+			throw new InvalidOperationException("Failed to create exception handler trampoline");
+		}
+		_exceptionHandler = (nint)AddVectoredExceptionHandler(1u, _exceptionHandlerStub);
 		Console.Error.WriteLine($"[LOADER][INFO] Exception handler installed: 0x{_exceptionHandler:X16}");
 
 		_unhandledFilterDelegate = UnhandledExceptionFilter;
 		_unhandledFilterHandle = GCHandle.Alloc(_unhandledFilterDelegate);
-		SetUnhandledExceptionFilter(Marshal.GetFunctionPointerForDelegate(_unhandledFilterDelegate));
+		_unhandledFilterStub = CreateExceptionHandlerTrampoline(Marshal.GetFunctionPointerForDelegate(_unhandledFilterDelegate));
+		if (_unhandledFilterStub == 0)
+		{
+			throw new InvalidOperationException("Failed to create unhandled exception filter trampoline");
+		}
+		SetUnhandledExceptionFilter(_unhandledFilterStub);
 	}
 
 	private unsafe int UnhandledExceptionFilter(void* exceptionInfo)

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

@@ -143,6 +143,8 @@ public sealed partial class DirectExecutionBackend
 			ProbeReturnRip(num7, num);
 		}
 		TrackStrlenPrelude(importStubEntry.Nid, num, num7);
+		TraceGuestContext(
+			$"import dispatch={num} nid={importStubEntry.Nid} ret=0x{num7:X16} managed={Environment.CurrentManagedThreadId} guest=0x{GuestThreadExecution.CurrentGuestThreadHandle:X16} fiber=0x{GuestThreadExecution.CurrentFiberAddress:X16} active={HasActiveExecutionThread}");
 		bool logBootstrap = string.Equals(Environment.GetEnvironmentVariable("SHARPEMU_LOG_BOOTSTRAP"), "1", StringComparison.Ordinal);
 		if (logBootstrap && string.Equals(importStubEntry.Nid, RuntimeStubNids.BootstrapBridge, StringComparison.Ordinal))
 		{
@@ -279,13 +281,15 @@ public sealed partial class DirectExecutionBackend
 			catch
 			{
 			}
-			TryBypassStackChkFailTrap(num, num7);
 		}
 		try
 		{
 			OrbisGen2Result orbisGen2Result;
 			bool dispatchResolved = true;
-			var previousImportCallFrame = GuestThreadExecution.EnterImportCallFrame(num7, (ulong)argPackPtr + 104uL);
+			var previousImportCallFrame = GuestThreadExecution.EnterImportCallFrame(
+				num7,
+				(ulong)argPackPtr + 104uL,
+				ActiveGuestReturnSlotAddress);
 			try
 			{
 				if (string.Equals(importStubEntry.Nid, RuntimeStubNids.BootstrapBridge, StringComparison.Ordinal))
@@ -408,7 +412,7 @@ public sealed partial class DirectExecutionBackend
 	private unsafe bool TryForceGuestExitToHostStub(nint argPackPtr, long dispatchIndex, ulong returnRip, string nid)
 	{
 		ulong num = ActiveEntryReturnSentinelRip;
-		if (num < 65536)
+		if (num < 65536 || !TryPatchActiveGuestReturnSlot(num))
 		{
 			return false;
 		}
@@ -430,7 +434,7 @@ public sealed partial class DirectExecutionBackend
 	private unsafe bool TryCompleteGuestEntryToHostStub(nint argPackPtr, long dispatchIndex, ulong returnRip, string nid, string reason, int status)
 	{
 		ulong hostExit = ActiveEntryReturnSentinelRip;
-		if (hostExit < 65536)
+		if (hostExit < 65536 || !TryPatchActiveGuestReturnSlot(hostExit))
 		{
 			return false;
 		}
@@ -450,7 +454,7 @@ public sealed partial class DirectExecutionBackend
 	private unsafe bool TryYieldGuestThreadToHostStub(nint argPackPtr, long dispatchIndex, ulong returnRip, string nid, string reason)
 	{
 		ulong hostExit = ActiveEntryReturnSentinelRip;
-		if (hostExit < 65536)
+		if (hostExit < 65536 || !TryPatchActiveGuestReturnSlot(hostExit))
 		{
 			return false;
 		}
@@ -470,6 +474,14 @@ public sealed partial class DirectExecutionBackend
 		return true;
 	}
 
+	private bool TryPatchActiveGuestReturnSlot(ulong hostExit)
+	{
+		ulong returnSlotAddress = ActiveGuestReturnSlotAddress;
+		return returnSlotAddress != 0 &&
+			ActiveCpuContext is not null &&
+			ActiveCpuContext.TryWriteUInt64(returnSlotAddress, hostExit);
+	}
+
 	private bool ShouldForceGuestExitOnImportLoop(string nid, ulong returnRip, long dispatchIndex, ulong arg0, ulong arg1)
 	{
 		if (dispatchIndex < 1200)
@@ -480,6 +492,11 @@ public sealed partial class DirectExecutionBackend
 		{
 			return false;
 		}
+		if (IsImportLoopGuardBoundary(nid))
+		{
+			ResetImportLoopPattern();
+			return false;
+		}
 		if (!_importNidHashCache.TryGetValue(nid, out var value))
 		{
 			value = StableHash64(nid);
@@ -513,6 +530,17 @@ public sealed partial class DirectExecutionBackend
 		return elapsedTicks >= (long)(guardSeconds * Stopwatch.Frequency);
 	}
 
+	private static bool IsImportLoopGuardBoundary(string nid) =>
+		string.Equals(nid, "1jfXLRVzisc", StringComparison.Ordinal);
+
+	private void ResetImportLoopPattern()
+	{
+		_importLoopPatternHits = 0;
+		_importLoopPatternStartTimestamp = 0;
+		_importLoopSignatureCount = 0;
+		_importLoopSignatureWriteIndex = 0;
+	}
+
 	private static int GetImportLoopGuardSeconds()
 	{
 		if (int.TryParse(Environment.GetEnvironmentVariable("SHARPEMU_IMPORT_LOOP_GUARD_SECONDS"), out var seconds))
@@ -923,66 +951,6 @@ public sealed partial class DirectExecutionBackend
 		}
 	}
 
-	private void TryBypassStackChkFailTrap(long dispatchIndex, ulong returnRip)
-	{
-		var cpuContext = ActiveCpuContext;
-		if (cpuContext == null || returnRip < 32)
-		{
-			return;
-		}
-		try
-		{
-			byte[] array = new byte[19];
-			ulong num = returnRip - 23;
-			Marshal.Copy((nint)num, array, 0, array.Length);
-			if (array[0] != 117 || array[1] != 16 || array[2] != 72 || array[3] != 137 || array[4] != 216 || array[5] != 72 || array[6] != 131 || array[7] != 196 || array[9] != 91 || array[10] != 65 || array[11] != 92 || array[12] != 65 || array[13] != 94 || array[14] != 65 || array[15] != 95 || array[16] != 93 || array[17] != 195 || array[18] != 232)
-			{
-				return;
-			}
-			ulong value = returnRip - 21;
-			ulong address = cpuContext[CpuRegister.Rsp];
-			if (cpuContext.TryWriteUInt64(address, value))
-			{
-				if (_stackChkBypassSites.Add(num) && TryPatchStackChkFailBranch(num))
-				{
-					Console.Error.WriteLine($"[LOADER][WARNING] Import#{dispatchIndex}: patched stack_chk_fail tail branch at 0x{num:X16} -> NOP NOP");
-				}
-				Console.Error.WriteLine($"[LOADER][WARNING] Import#{dispatchIndex}: redirected __stack_chk_fail return to epilogue 0x{value:X16}");
-			}
-		}
-		catch
-		{
-		}
-	}
-
-	private unsafe static bool TryPatchStackChkFailBranch(ulong branchAddress)
-	{
-		uint flNewProtect = default(uint);
-		if (!VirtualProtect((void*)branchAddress, 2u, 64u, &flNewProtect))
-		{
-			return false;
-		}
-		try
-		{
-			if (Marshal.ReadByte((nint)branchAddress) != 117)
-			{
-				return false;
-			}
-			Marshal.WriteByte((nint)branchAddress, 144);
-			Marshal.WriteByte((nint)(branchAddress + 1), 144);
-			FlushInstructionCache(GetCurrentProcess(), (void*)branchAddress, 2u);
-			return true;
-		}
-		catch
-		{
-			return false;
-		}
-		finally
-		{
-			VirtualProtect((void*)branchAddress, 2u, flNewProtect, &flNewProtect);
-		}
-	}
-
 	private unsafe void TryPatchEa020eLookupCall(long dispatchIndex, ulong returnRip)
 	{
 		if (_patchedEa020eLookupCall || returnRip != 0x0000000800EA01A6uL)

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

@@ -136,10 +136,18 @@ public sealed unsafe partial class DirectExecutionBackend : INativeCpuBackend, I
 
 	private nint _unresolvedReturnStub;
 
+	private nint _guestReturnStub;
+
 	private nint _rawExceptionHandler;
 
+	private nint _rawExceptionHandlerStub;
+
 	private nint _exceptionHandler;
 
+	private nint _exceptionHandlerStub;
+
+	private nint _unhandledFilterStub;
+
 	private nint _lowIndexedTableScratch;
 
 	private nint _stackGuardCompareScratch;
@@ -161,6 +169,9 @@ public sealed unsafe partial class DirectExecutionBackend : INativeCpuBackend, I
 	[ThreadStatic]
 	private static ulong _activeEntryReturnSentinelRip;
 
+	[ThreadStatic]
+	private static ulong _activeGuestReturnSlotAddress;
+
 	[ThreadStatic]
 	private static bool _activeForcedGuestExit;
 
@@ -202,8 +213,6 @@ public sealed unsafe partial class DirectExecutionBackend : INativeCpuBackend, I
 
 	private readonly HashSet<ulong> _fallbackTrapStubs = new HashSet<ulong>();
 
-	private readonly HashSet<ulong> _stackChkBypassSites = new HashSet<ulong>();
-
 	private readonly HashSet<ulong> _patchedResolverReturnSites = new HashSet<ulong>();
 
 	private readonly HashSet<ulong> _patchedTlsImmediateThunkTargets = new HashSet<ulong>();
@@ -539,6 +548,9 @@ public sealed unsafe partial class DirectExecutionBackend : INativeCpuBackend, I
 		}
 	}
 
+	private ulong ActiveGuestReturnSlotAddress =>
+		HasActiveExecutionThread ? _activeGuestReturnSlotAddress : 0;
+
 	private bool ActiveForcedGuestExit
 	{
 		get => HasActiveExecutionThread ? _activeForcedGuestExit : _forcedGuestExit;
@@ -591,6 +603,7 @@ public sealed unsafe partial class DirectExecutionBackend : INativeCpuBackend, I
 		DirectExecutionBackend? previousBackend,
 		CpuContext? previousContext,
 		ulong previousSentinel,
+		ulong previousReturnSlotAddress,
 		bool previousForcedExit,
 		bool previousYieldRequested,
 		string? previousYieldReason)
@@ -598,6 +611,7 @@ public sealed unsafe partial class DirectExecutionBackend : INativeCpuBackend, I
 		_activeExecutionBackend = previousBackend;
 		_activeCpuContext = previousContext;
 		_activeEntryReturnSentinelRip = previousSentinel;
+		_activeGuestReturnSlotAddress = previousReturnSlotAddress;
 		_activeForcedGuestExit = previousForcedExit;
 		_activeGuestThreadYieldRequested = previousYieldRequested;
 		_activeGuestThreadYieldReason = previousYieldReason;
@@ -634,6 +648,11 @@ public sealed unsafe partial class DirectExecutionBackend : INativeCpuBackend, I
 			throw new OutOfMemoryException("Failed to allocate host stack slot storage");
 		}
 		_unresolvedReturnStub = CreateUnresolvedReturnStub();
+		_guestReturnStub = CreateGuestReturnStub();
+		if (_guestReturnStub == 0)
+		{
+			throw new OutOfMemoryException("Failed to allocate guest return stub");
+		}
 		SetupExceptionHandler();
 	}
 
@@ -930,9 +949,24 @@ public sealed unsafe partial class DirectExecutionBackend : INativeCpuBackend, I
 		yield return "_" + exportName;
 	}
 
-	private static bool IsDirectImportTargetUsable(ulong address)
+	private bool IsDirectImportTargetUsable(ulong address)
 	{
-		return address >= 65536 && !IsUnresolvedSentinel(address);
+		if (address < 65536 || IsUnresolvedSentinel(address) ||
+			_cpuContext is null || !TryGetVirtualMemory(_cpuContext, out var virtualMemory))
+		{
+			return false;
+		}
+
+		foreach (var region in virtualMemory.SnapshotRegions())
+		{
+			if ((region.Protection & ProgramHeaderFlags.Execute) != 0 &&
+				ContainsAddress(region.VirtualAddress, region.MemorySize, address))
+			{
+				return true;
+			}
+		}
+
+		return false;
 	}
 
 	private unsafe void BindTlsBase(CpuContext context)
@@ -1243,6 +1277,145 @@ public sealed unsafe partial class DirectExecutionBackend : INativeCpuBackend, I
 		return (nint)ptr;
 	}
 
+	private unsafe nint CreateGuestReturnStub()
+	{
+		const uint stubSize = 256u;
+		void* ptr = VirtualAlloc(null, stubSize, 12288u, 64u);
+		if (ptr == null)
+		{
+			return 0;
+		}
+
+		byte* code = (byte*)ptr;
+		int offset = 0;
+		EmitByte(code, ref offset, 0x48); // sub rsp, 0x20
+		EmitByte(code, ref offset, 0x83);
+		EmitByte(code, ref offset, 0xEC);
+		EmitByte(code, ref offset, 0x20);
+		EmitByte(code, ref offset, 0xB9); // mov ecx, tlsIndex
+		EmitUInt32(code, ref offset, _hostRspSlotTlsIndex);
+		EmitByte(code, ref offset, 0x48); // mov rax, TlsGetValue
+		EmitByte(code, ref offset, 0xB8);
+		*(long*)(code + offset) = _tlsGetValueAddress;
+		offset += sizeof(ulong);
+		EmitByte(code, ref offset, 0xFF); // call rax
+		EmitByte(code, ref offset, 0xD0);
+		EmitByte(code, ref offset, 0x48); // add rsp, 0x20
+		EmitByte(code, ref offset, 0x83);
+		EmitByte(code, ref offset, 0xC4);
+		EmitByte(code, ref offset, 0x20);
+		EmitByte(code, ref offset, 0x48); // mov rsp, [rax]
+		EmitByte(code, ref offset, 0x8B);
+		EmitByte(code, ref offset, 0x20);
+		EmitHostNonvolatileXmmRestore(code, ref offset);
+		EmitByte(code, ref offset, 0x41); EmitByte(code, ref offset, 0x5F);
+		EmitByte(code, ref offset, 0x41); EmitByte(code, ref offset, 0x5E);
+		EmitByte(code, ref offset, 0x41); EmitByte(code, ref offset, 0x5D);
+		EmitByte(code, ref offset, 0x41); EmitByte(code, ref offset, 0x5C);
+		EmitByte(code, ref offset, 0x5E);
+		EmitByte(code, ref offset, 0x5F);
+		EmitByte(code, ref offset, 0x5D);
+		EmitByte(code, ref offset, 0x5B);
+		EmitByte(code, ref offset, 0xC3);
+
+		uint oldProtect = default;
+		VirtualProtect(ptr, stubSize, 32u, &oldProtect);
+		FlushInstructionCache(GetCurrentProcess(), ptr, (nuint)offset);
+		return (nint)ptr;
+	}
+
+	private unsafe nint CreateExceptionHandlerTrampoline(nint managedHandler)
+	{
+		const uint stubSize = 256u;
+		void* ptr = VirtualAlloc(null, stubSize, 12288u, 64u);
+		if (ptr == null)
+		{
+			return 0;
+		}
+
+		byte* code = (byte*)ptr;
+		int offset = 0;
+		EmitByte(code, ref offset, 0x41); EmitByte(code, ref offset, 0x54); // push r12
+		EmitByte(code, ref offset, 0x41); EmitByte(code, ref offset, 0x55); // push r13
+		EmitByte(code, ref offset, 0x49); EmitByte(code, ref offset, 0x89); EmitByte(code, ref offset, 0xE4); // mov r12, rsp
+		EmitByte(code, ref offset, 0x49); EmitByte(code, ref offset, 0x89); EmitByte(code, ref offset, 0xCD); // mov r13, rcx
+		EmitByte(code, ref offset, 0x65); EmitByte(code, ref offset, 0x48); // mov rax, gs:[8]
+		EmitByte(code, ref offset, 0x8B); EmitByte(code, ref offset, 0x04); EmitByte(code, ref offset, 0x25);
+		EmitUInt32(code, ref offset, 8u);
+		EmitByte(code, ref offset, 0x49); EmitByte(code, ref offset, 0x39); EmitByte(code, ref offset, 0xC4); // cmp r12, rax
+		EmitByte(code, ref offset, 0x0F); EmitByte(code, ref offset, 0x83); // jae guestStack
+		int aboveStackJump = offset;
+		EmitUInt32(code, ref offset, 0u);
+		EmitByte(code, ref offset, 0x65); EmitByte(code, ref offset, 0x48); // mov rax, gs:[0x10]
+		EmitByte(code, ref offset, 0x8B); EmitByte(code, ref offset, 0x04); EmitByte(code, ref offset, 0x25);
+		EmitUInt32(code, ref offset, 0x10u);
+		EmitByte(code, ref offset, 0x49); EmitByte(code, ref offset, 0x39); EmitByte(code, ref offset, 0xC4); // cmp r12, rax
+		EmitByte(code, ref offset, 0x0F); EmitByte(code, ref offset, 0x82); // jb guestStack
+		int belowStackJump = offset;
+		EmitUInt32(code, ref offset, 0u);
+
+		EmitByte(code, ref offset, 0x48); EmitByte(code, ref offset, 0x83); EmitByte(code, ref offset, 0xEC); EmitByte(code, ref offset, 0x28);
+		EmitByte(code, ref offset, 0x4C); EmitByte(code, ref offset, 0x89); EmitByte(code, ref offset, 0xE9); // mov rcx, r13
+		EmitByte(code, ref offset, 0x48); EmitByte(code, ref offset, 0xB8);
+		*(nint*)(code + offset) = managedHandler;
+		offset += sizeof(nint);
+		EmitByte(code, ref offset, 0xFF); EmitByte(code, ref offset, 0xD0);
+		EmitByte(code, ref offset, 0x48); EmitByte(code, ref offset, 0x83); EmitByte(code, ref offset, 0xC4); EmitByte(code, ref offset, 0x28);
+		EmitByte(code, ref offset, 0xE9);
+		int hostRestoreJump = offset;
+		EmitUInt32(code, ref offset, 0u);
+
+		int guestStackOffset = offset;
+		EmitByte(code, ref offset, 0x48); EmitByte(code, ref offset, 0x83); EmitByte(code, ref offset, 0xEC); EmitByte(code, ref offset, 0x28);
+		EmitByte(code, ref offset, 0xB9);
+		EmitUInt32(code, ref offset, _hostRspSlotTlsIndex);
+		EmitByte(code, ref offset, 0x48); EmitByte(code, ref offset, 0xB8);
+		*(nint*)(code + offset) = _tlsGetValueAddress;
+		offset += sizeof(nint);
+		EmitByte(code, ref offset, 0xFF); EmitByte(code, ref offset, 0xD0);
+		EmitByte(code, ref offset, 0x48); EmitByte(code, ref offset, 0x83); EmitByte(code, ref offset, 0xC4); EmitByte(code, ref offset, 0x28);
+		EmitByte(code, ref offset, 0x48); EmitByte(code, ref offset, 0x85); EmitByte(code, ref offset, 0xC0); // test rax, rax
+		EmitByte(code, ref offset, 0x0F); EmitByte(code, ref offset, 0x84);
+		int missingTlsJump = offset;
+		EmitUInt32(code, ref offset, 0u);
+		EmitByte(code, ref offset, 0x4C); EmitByte(code, ref offset, 0x8B); EmitByte(code, ref offset, 0x18); // mov r11, [rax]
+		EmitByte(code, ref offset, 0x4D); EmitByte(code, ref offset, 0x85); EmitByte(code, ref offset, 0xDB); // test r11, r11
+		EmitByte(code, ref offset, 0x0F); EmitByte(code, ref offset, 0x84);
+		int missingHostStackJump = offset;
+		EmitUInt32(code, ref offset, 0u);
+		EmitByte(code, ref offset, 0x4C); EmitByte(code, ref offset, 0x89); EmitByte(code, ref offset, 0xDC); // mov rsp, r11
+		EmitByte(code, ref offset, 0x48); EmitByte(code, ref offset, 0x83); EmitByte(code, ref offset, 0xEC); EmitByte(code, ref offset, 0x28);
+		EmitByte(code, ref offset, 0x4C); EmitByte(code, ref offset, 0x89); EmitByte(code, ref offset, 0xE9); // mov rcx, r13
+		EmitByte(code, ref offset, 0x48); EmitByte(code, ref offset, 0xB8);
+		*(nint*)(code + offset) = managedHandler;
+		offset += sizeof(nint);
+		EmitByte(code, ref offset, 0xFF); EmitByte(code, ref offset, 0xD0);
+		EmitByte(code, ref offset, 0x48); EmitByte(code, ref offset, 0x83); EmitByte(code, ref offset, 0xC4); EmitByte(code, ref offset, 0x28);
+		EmitByte(code, ref offset, 0xE9);
+		int guestRestoreJump = offset;
+		EmitUInt32(code, ref offset, 0u);
+
+		int passThroughOffset = offset;
+		EmitByte(code, ref offset, 0x31); EmitByte(code, ref offset, 0xC0); // xor eax, eax
+		int restoreOffset = offset;
+		EmitByte(code, ref offset, 0x4C); EmitByte(code, ref offset, 0x89); EmitByte(code, ref offset, 0xE4); // mov rsp, r12
+		EmitByte(code, ref offset, 0x41); EmitByte(code, ref offset, 0x5D);
+		EmitByte(code, ref offset, 0x41); EmitByte(code, ref offset, 0x5C);
+		EmitByte(code, ref offset, 0xC3);
+
+		*(int*)(code + aboveStackJump) = guestStackOffset - (aboveStackJump + sizeof(int));
+		*(int*)(code + belowStackJump) = guestStackOffset - (belowStackJump + sizeof(int));
+		*(int*)(code + hostRestoreJump) = restoreOffset - (hostRestoreJump + sizeof(int));
+		*(int*)(code + missingTlsJump) = passThroughOffset - (missingTlsJump + sizeof(int));
+		*(int*)(code + missingHostStackJump) = passThroughOffset - (missingHostStackJump + sizeof(int));
+		*(int*)(code + guestRestoreJump) = restoreOffset - (guestRestoreJump + sizeof(int));
+
+		uint oldProtect = default;
+		VirtualProtect(ptr, stubSize, 32u, &oldProtect);
+		FlushInstructionCache(GetCurrentProcess(), ptr, (nuint)offset);
+		return (nint)ptr;
+	}
+
 	private unsafe void* TryAllocateNearEntry(nuint size)
 	{
 		ulong entryPoint = _entryPoint;
@@ -1980,8 +2153,8 @@ public sealed unsafe partial class DirectExecutionBackend : INativeCpuBackend, I
 		{
 			Rip = continuation.Rip,
 			Rflags = continuation.Rflags == 0 ? 0x202UL : continuation.Rflags,
-			FsBase = continuation.FsBase != 0 ? continuation.FsBase : (callerContext.FsBase != 0 ? callerContext.FsBase : fallbackTlsBase),
+			FsBase = callerContext.FsBase != 0 ? callerContext.FsBase : (continuation.FsBase != 0 ? continuation.FsBase : fallbackTlsBase),
-			GsBase = continuation.GsBase != 0 ? continuation.GsBase : (callerContext.GsBase != 0 ? callerContext.GsBase : fallbackTlsBase),
+			GsBase = callerContext.GsBase != 0 ? callerContext.GsBase : (continuation.GsBase != 0 ? continuation.GsBase : fallbackTlsBase),
 		};
 
 		context[CpuRegister.Rax] = continuation.Rax;
@@ -1997,21 +2170,39 @@ public sealed unsafe partial class DirectExecutionBackend : INativeCpuBackend, I
 		context[CpuRegister.R13] = continuation.R13;
 		context[CpuRegister.R14] = continuation.R14;
 		context[CpuRegister.R15] = continuation.R15;
-		context[CpuRegister.Rsp] = continuation.Rsp + 16uL;
+		context[CpuRegister.Rsp] = continuation.Rsp;
 
-		var currentGuestThreadHandle = GuestThreadExecution.CurrentGuestThreadHandle;
 		var exitReason = GuestNativeCallExitReason.Exception;
 		string? callbackReason = null;
 		string? callbackLastError = null;
 		Exception? callbackException = null;
+		var currentGuestThreadHandle = GuestThreadExecution.CurrentGuestThreadHandle;
+		var currentFiberAddress = GuestThreadExecution.CurrentFiberAddress;
 
 		void RunContinuation()
 		{
+			var restoreGuestThread = currentGuestThreadHandle != 0 &&
+				GuestThreadExecution.CurrentGuestThreadHandle != currentGuestThreadHandle;
+			var previousGuestThreadHandle = restoreGuestThread
+				? GuestThreadExecution.EnterGuestThread(currentGuestThreadHandle)
+				: 0UL;
+			var restoreFiber = currentFiberAddress != 0 &&
+				GuestThreadExecution.CurrentFiberAddress != currentFiberAddress;
+			var previousFiberAddress = restoreFiber
+				? GuestThreadExecution.EnterFiber(currentFiberAddress)
+				: 0UL;
 			var previousLastError = LastError;
 			try
 			{
+				TraceGuestContext(
+					$"continuation-enter reason={reason} managed={Environment.CurrentManagedThreadId} guest=0x{GuestThreadExecution.CurrentGuestThreadHandle:X16} fiber=0x{GuestThreadExecution.CurrentFiberAddress:X16} captured_guest=0x{currentGuestThreadHandle:X16} captured_fiber=0x{currentFiberAddress:X16} restore_guest={restoreGuestThread} restore_fiber={restoreFiber}");
 				LastError = null;
-				exitReason = ExecuteGuestContinuationEntry(context, continuation.Rip, reason, out callbackReason);
+				exitReason = ExecuteGuestContinuationEntry(
+					context,
+					continuation.Rip,
+					continuation.ReturnSlotAddress,
+					reason,
+					out callbackReason);
 				callbackLastError = LastError;
 			}
 			catch (Exception ex)
@@ -2022,7 +2213,17 @@ public sealed unsafe partial class DirectExecutionBackend : INativeCpuBackend, I
 			}
 			finally
 			{
+				TraceGuestContext(
+					$"continuation-exit reason={reason} managed={Environment.CurrentManagedThreadId} guest=0x{GuestThreadExecution.CurrentGuestThreadHandle:X16} fiber=0x{GuestThreadExecution.CurrentFiberAddress:X16} exit={exitReason}");
 				LastError = previousLastError;
+				if (restoreFiber)
+				{
+					GuestThreadExecution.RestoreFiber(previousFiberAddress);
+				}
+				if (restoreGuestThread)
+				{
+					GuestThreadExecution.RestoreGuestThread(previousGuestThreadHandle);
+				}
 			}
 		}
 
@@ -2045,6 +2246,12 @@ public sealed unsafe partial class DirectExecutionBackend : INativeCpuBackend, I
 			{
 				runner.Run(RunContinuation);
 			}
+			else if (runner is not null)
+			{
+				TraceGuestContext(
+					$"continuation-inline reason={reason} managed={Environment.CurrentManagedThreadId} guest=0x{currentGuestThreadHandle:X16} fiber=0x{currentFiberAddress:X16}");
+				RunContinuation();
+			}
 			else
 			{
 				RunContinuationOnTemporaryThread(currentGuestThreadHandle, RunContinuation);
@@ -2070,6 +2277,14 @@ public sealed unsafe partial class DirectExecutionBackend : INativeCpuBackend, I
 		return true;
 	}
 
+	private static void TraceGuestContext(string message)
+	{
+		if (string.Equals(Environment.GetEnvironmentVariable("SHARPEMU_LOG_CONTEXT"), "1", StringComparison.Ordinal))
+		{
+			Console.Error.WriteLine($"[LOADER][TRACE] guest_context.{message}");
+		}
+	}
+
 	private static void RunContinuationOnTemporaryThread(ulong guestThreadHandle, Action continuation)
 	{
 		var continuationThread = new Thread(() =>
@@ -2351,6 +2566,7 @@ public sealed unsafe partial class DirectExecutionBackend : INativeCpuBackend, I
 		var previousActiveBackend = _activeExecutionBackend;
 		var previousActiveContext = _activeCpuContext;
 		var previousSentinel = _activeEntryReturnSentinelRip;
+		var previousReturnSlotAddress = _activeGuestReturnSlotAddress;
 		var previousForcedExit = _activeForcedGuestExit;
 		var previousYieldRequested = _activeGuestThreadYieldRequested;
 		var previousYieldReason = _activeGuestThreadYieldReason;
@@ -2360,6 +2576,7 @@ public sealed unsafe partial class DirectExecutionBackend : INativeCpuBackend, I
 			_activeExecutionBackend = this;
 			_activeCpuContext = context;
 			_activeEntryReturnSentinelRip = 0;
+			_activeGuestReturnSlotAddress = 0;
 			_activeForcedGuestExit = false;
 			_activeGuestThreadYieldRequested = false;
 			_activeGuestThreadYieldReason = null;
@@ -2463,7 +2680,8 @@ public sealed unsafe partial class DirectExecutionBackend : INativeCpuBackend, I
 			ptr2[offset++] = 91;
 			ptr2[offset++] = 195;
 			ulong sentinel = (ulong)ptr + (ulong)sentinelOffset;
-			ActiveEntryReturnSentinelRip = sentinel;
+			ActiveEntryReturnSentinelRip = (ulong)_guestReturnStub;
+			_activeGuestReturnSlotAddress = context[CpuRegister.Rsp] - 16uL;
 			if (!context.TryWriteUInt64(context[CpuRegister.Rsp], sentinel))
 			{
 				reason = $"failed to patch guest thread return sentinel at 0x{context[CpuRegister.Rsp]:X16}";
@@ -2509,6 +2727,7 @@ public sealed unsafe partial class DirectExecutionBackend : INativeCpuBackend, I
 				previousActiveBackend,
 				previousActiveContext,
 				previousSentinel,
+				previousReturnSlotAddress,
 				previousForcedExit,
 				previousYieldRequested,
 				previousYieldReason);
@@ -2516,7 +2735,12 @@ public sealed unsafe partial class DirectExecutionBackend : INativeCpuBackend, I
 		}
 	}
 
-	private unsafe GuestNativeCallExitReason ExecuteGuestContinuationEntry(CpuContext context, ulong entryPoint, string name, out string? reason)
+	private unsafe GuestNativeCallExitReason ExecuteGuestContinuationEntry(
+		CpuContext context,
+		ulong entryPoint,
+		ulong returnSlotAddress,
+		string name,
+		out string? reason)
 	{
 		reason = null;
 		if (context[CpuRegister.Rsp] == 0)
@@ -2534,6 +2758,7 @@ public sealed unsafe partial class DirectExecutionBackend : INativeCpuBackend, I
 		var previousActiveBackend = _activeExecutionBackend;
 		var previousActiveContext = _activeCpuContext;
 		var previousSentinel = _activeEntryReturnSentinelRip;
+		var previousReturnSlotAddress = _activeGuestReturnSlotAddress;
 		var previousForcedExit = _activeForcedGuestExit;
 		var previousYieldRequested = _activeGuestThreadYieldRequested;
 		var previousYieldReason = _activeGuestThreadYieldReason;
@@ -2543,6 +2768,7 @@ public sealed unsafe partial class DirectExecutionBackend : INativeCpuBackend, I
 			_activeExecutionBackend = this;
 			_activeCpuContext = context;
 			_activeEntryReturnSentinelRip = 0;
+			_activeGuestReturnSlotAddress = returnSlotAddress;
 			_activeForcedGuestExit = false;
 			_activeGuestThreadYieldRequested = false;
 			_activeGuestThreadYieldReason = null;
@@ -2577,7 +2803,6 @@ public sealed unsafe partial class DirectExecutionBackend : INativeCpuBackend, I
 			Emit(0x49); Emit(0x89); Emit(0x22); // mov [r10], rsp
 			EmitMovR64Imm(0x48, 0xB8, context[CpuRegister.Rsp]); // mov rax, guest rsp
 			Emit(0x48); Emit(0x89); Emit(0xC4); // mov rsp, rax
-			Emit(0x48); Emit(0x83); Emit(0xEC); Emit(0x08); // sub rsp, 8
 			EmitMovR64Imm(0x48, 0xBB, context[CpuRegister.Rbx]); // mov rbx, imm64
 			EmitMovR64Imm(0x48, 0xBD, context[CpuRegister.Rbp]); // mov rbp, imm64
 			EmitMovR64Imm(0x48, 0xBF, context[CpuRegister.Rdi]); // mov rdi, imm64
@@ -2592,29 +2817,11 @@ public sealed unsafe partial class DirectExecutionBackend : INativeCpuBackend, I
 			EmitMovR64Imm(0x49, 0xBF, context[CpuRegister.R15]); // mov r15, imm64
 			EmitMovR64Imm(0x48, 0xB8, context[CpuRegister.Rax]); // mov rax, imm64
 			EmitMovR64Imm(0x49, 0xBB, entryPoint); // mov r11, entryPoint
-			Emit(0x41); Emit(0xFF); Emit(0xD3); // call r11
+			Emit(0x41); Emit(0xFF); Emit(0xE3); // jmp r11
-			int sentinelOffset = offset + 4;
+			ActiveEntryReturnSentinelRip = (ulong)_guestReturnStub;
-			Emit(0x48); Emit(0x83); Emit(0xC4); Emit(0x08); // add rsp, 8
+			if (returnSlotAddress == 0 || !context.TryWriteUInt64(returnSlotAddress, (ulong)_guestReturnStub))
-			EmitMovR64Imm(0x49, 0xBA, hostRspSlot); // mov r10, hostRspSlot
-			Emit(0x49); Emit(0x8B); Emit(0x22); // mov rsp, [r10]
-			EmitHostNonvolatileXmmRestore(ptr2, ref offset);
-			Emit(0x41); Emit(0x5F); // pop r15
-			Emit(0x41); Emit(0x5E); // pop r14
-			Emit(0x41); Emit(0x5D); // pop r13
-			Emit(0x41); Emit(0x5C); // pop r12
-			Emit(0x5E); // pop rsi
-			Emit(0x5F); // pop rdi
-			Emit(0x5D); // pop rbp
-			Emit(0x5B); // pop rbx
-			Emit(0xC3); // ret
-			ulong sentinel = (ulong)ptr + (ulong)sentinelOffset;
-			ActiveEntryReturnSentinelRip = sentinel;
-			var sentinelStackAddress = context[CpuRegister.Rsp] >= 16uL
-				? context[CpuRegister.Rsp] - 16uL
-				: 0;
-			if (sentinelStackAddress == 0 || !context.TryWriteUInt64(sentinelStackAddress, sentinel))
 			{
-				reason = $"failed to patch guest continuation return sentinel at 0x{sentinelStackAddress:X16}";
+				reason = $"failed to patch guest continuation return slot at 0x{returnSlotAddress:X16}";
 				return GuestNativeCallExitReason.Exception;
 			}
 			uint oldProtect = default(uint);
@@ -2657,6 +2864,7 @@ public sealed unsafe partial class DirectExecutionBackend : INativeCpuBackend, I
 				previousActiveBackend,
 				previousActiveContext,
 				previousSentinel,
+				previousReturnSlotAddress,
 				previousForcedExit,
 				previousYieldRequested,
 				previousYieldReason);
@@ -2717,9 +2925,18 @@ public sealed unsafe partial class DirectExecutionBackend : INativeCpuBackend, I
 		}
 		EmitByte(code, ref offset, 0x0F);
 		EmitByte(code, ref offset, store ? (byte)0x7F : (byte)0x6F);
-		EmitByte(code, ref offset, (byte)(0x44 | ((xmm & 7) << 3)));
+		if (displacement < 0x80)
-		EmitByte(code, ref offset, 0x24);
+		{
-		EmitByte(code, ref offset, displacement);
+			EmitByte(code, ref offset, (byte)(0x44 | ((xmm & 7) << 3)));
+			EmitByte(code, ref offset, 0x24);
+			EmitByte(code, ref offset, displacement);
+		}
+		else
+		{
+			EmitByte(code, ref offset, (byte)(0x84 | ((xmm & 7) << 3)));
+			EmitByte(code, ref offset, 0x24);
+			EmitUInt32(code, ref offset, displacement);
+		}
 	}
 
 	private unsafe bool ExecuteEntry(CpuContext context, ulong entryPoint, out OrbisGen2Result result)
@@ -2745,6 +2962,7 @@ public sealed unsafe partial class DirectExecutionBackend : INativeCpuBackend, I
 		var previousActiveBackend = _activeExecutionBackend;
 		var previousActiveContext = _activeCpuContext;
 		var previousSentinel = _activeEntryReturnSentinelRip;
+		var previousReturnSlotAddress = _activeGuestReturnSlotAddress;
 		var previousForcedExit = _activeForcedGuestExit;
 		var previousYieldRequested = _activeGuestThreadYieldRequested;
 		var previousYieldReason = _activeGuestThreadYieldReason;
@@ -2754,6 +2972,7 @@ public sealed unsafe partial class DirectExecutionBackend : INativeCpuBackend, I
 			_activeExecutionBackend = this;
 			_activeCpuContext = context;
 			_activeEntryReturnSentinelRip = 0;
+			_activeGuestReturnSlotAddress = 0;
 			_activeForcedGuestExit = false;
 			_activeGuestThreadYieldRequested = false;
 			_activeGuestThreadYieldReason = null;
@@ -2857,7 +3076,8 @@ public sealed unsafe partial class DirectExecutionBackend : INativeCpuBackend, I
 			ptr2[num3++] = 91;
 			ptr2[num3++] = 195;
 			ulong value = (ulong)ptr + (ulong)num4;
-			ActiveEntryReturnSentinelRip = value;
+			ActiveEntryReturnSentinelRip = (ulong)_guestReturnStub;
+			_activeGuestReturnSlotAddress = context[CpuRegister.Rsp] - 16uL;
 			if (!context.TryWriteUInt64(context[CpuRegister.Rsp], value))
 			{
 				LastError = $"Failed to patch native return sentinel at 0x{context[CpuRegister.Rsp]:X16}";
@@ -2939,6 +3159,7 @@ public sealed unsafe partial class DirectExecutionBackend : INativeCpuBackend, I
 				previousActiveBackend,
 				previousActiveContext,
 				previousSentinel,
+				previousReturnSlotAddress,
 				previousForcedExit,
 				previousYieldRequested,
 				previousYieldReason);
@@ -3116,6 +3337,22 @@ public sealed unsafe partial class DirectExecutionBackend : INativeCpuBackend, I
 			RemoveVectoredExceptionHandler((void*)_rawExceptionHandler);
 			_rawExceptionHandler = 0;
 		}
+		if (_rawExceptionHandlerStub != 0)
+		{
+			VirtualFree((void*)_rawExceptionHandlerStub, 0u, 32768u);
+			_rawExceptionHandlerStub = 0;
+		}
+		if (_exceptionHandlerStub != 0)
+		{
+			VirtualFree((void*)_exceptionHandlerStub, 0u, 32768u);
+			_exceptionHandlerStub = 0;
+		}
+		if (_unhandledFilterStub != 0)
+		{
+			SetUnhandledExceptionFilter(0);
+			VirtualFree((void*)_unhandledFilterStub, 0u, 32768u);
+			_unhandledFilterStub = 0;
+		}
 		if (_handlerHandle.IsAllocated)
 		{
 			_handlerHandle.Free();
@@ -3172,6 +3409,11 @@ public sealed unsafe partial class DirectExecutionBackend : INativeCpuBackend, I
 			VirtualFree((void*)_unresolvedReturnStub, 0u, 32768u);
 			_unresolvedReturnStub = 0;
 		}
+		if (_guestReturnStub != 0)
+		{
+			VirtualFree((void*)_guestReturnStub, 0u, 32768u);
+			_guestReturnStub = 0;
+		}
 		if (_lowIndexedTableScratch != 0)
 		{
 			VirtualFree((void*)_lowIndexedTableScratch, 0u, 32768u);

src/SharpEmu.Core/Memory/PhysicalVirtualMemory.cs

@@ -222,6 +222,56 @@ public sealed unsafe class PhysicalVirtualMemory : IVirtualMemory, IGuestMemoryA
         return actualAddress;
     }
 
+    public bool TryAllocateAtOrAbove(
+        ulong desiredAddress,
+        ulong size,
+        bool executable,
+        ulong alignment,
+        out ulong actualAddress)
+    {
+        actualAddress = 0;
+        if (size == 0)
+        {
+            return false;
+        }
+
+        var alignedSize = AlignUp(size, PageSize);
+        var effectiveAlignment = Math.Max(PageSize, alignment == 0 ? PageSize : alignment);
+        var cursor = AlignUp(desiredAddress, effectiveAlignment);
+
+        for (var attempt = 0; attempt < 0x10000; attempt++)
+        {
+            if (cursor == 0 || ulong.MaxValue - cursor < alignedSize)
+            {
+                return false;
+            }
+
+            if (TryGetOverlappingRegionEnd(cursor, alignedSize, out var overlapEnd))
+            {
+                cursor = AlignUp(overlapEnd, effectiveAlignment);
+                continue;
+            }
+
+            try
+            {
+                actualAddress = AllocateAt(cursor, alignedSize, executable, allowAlternative: false);
+                if (actualAddress == cursor)
+                {
+                    return true;
+                }
+
+                actualAddress = 0;
+            }
+            catch
+            {
+            }
+
+            cursor = AlignUp(cursor + effectiveAlignment, effectiveAlignment);
+        }
+
+        return false;
+    }
+
     public bool TryAllocateGuestMemory(ulong size, ulong alignment, out ulong address)
     {
         address = 0;
@@ -532,6 +582,30 @@ public sealed unsafe class PhysicalVirtualMemory : IVirtualMemory, IGuestMemoryA
         return null;
     }
 
+    private bool TryGetOverlappingRegionEnd(ulong address, ulong size, out ulong overlapEnd)
+    {
+        overlapEnd = 0;
+        if (size == 0 || ulong.MaxValue - address < size - 1)
+        {
+            return false;
+        }
+
+        var end = address + size;
+        lock (_gate)
+        {
+            foreach (var region in _regions)
+            {
+                var regionEnd = region.VirtualAddress + region.Size;
+                if (address < regionEnd && region.VirtualAddress < end)
+                {
+                    overlapEnd = Math.Max(overlapEnd, regionEnd);
+                }
+            }
+        }
+
+        return overlapEnd != 0;
+    }
+
     private static bool TryResolveRegionOffset(ulong address, ulong size, MemoryRegion region, out ulong offset)
     {
         offset = 0;

src/SharpEmu.HLE/GuestThreadExecution.cs

@@ -36,11 +36,13 @@ public interface IGuestThreadScheduler
 public readonly record struct GuestImportCallFrame(
     bool IsValid,
     ulong ReturnRip,
-    ulong ResumeRsp);
+    ulong ResumeRsp,
+    ulong ReturnSlotAddress);
 
 public readonly record struct GuestCpuContinuation(
     ulong Rip,
     ulong Rsp,
+    ulong ReturnSlotAddress,
     ulong Rflags,
     ulong FsBase,
     ulong GsBase,
@@ -63,6 +65,9 @@ public static class GuestThreadExecution
     [ThreadStatic]
     private static ulong _currentGuestThreadHandle;
 
+    [ThreadStatic]
+    private static ulong _currentFiberAddress;
+
     [ThreadStatic]
     private static string? _pendingBlockReason;
 
@@ -84,12 +89,17 @@ public static class GuestThreadExecution
     [ThreadStatic]
     private static ulong _currentImportResumeRsp;
 
+    [ThreadStatic]
+    private static ulong _currentImportReturnSlotAddress;
+
     public static IGuestThreadScheduler? Scheduler { get; set; }
 
     public static bool IsGuestThread => _currentGuestThreadHandle != 0;
 
     public static ulong CurrentGuestThreadHandle => _currentGuestThreadHandle;
 
+    public static ulong CurrentFiberAddress => _currentFiberAddress;
+
     public static ulong EnterGuestThread(ulong threadHandle)
     {
         var previous = _currentGuestThreadHandle;
@@ -101,6 +111,7 @@ public static class GuestThreadExecution
         _hasCurrentImportCallFrame = false;
         _currentImportReturnRip = 0;
         _currentImportResumeRsp = 0;
+        _currentImportReturnSlotAddress = 0;
         return previous;
     }
 
@@ -114,6 +125,19 @@ public static class GuestThreadExecution
         _hasCurrentImportCallFrame = false;
         _currentImportReturnRip = 0;
         _currentImportResumeRsp = 0;
+        _currentImportReturnSlotAddress = 0;
+    }
+
+    public static ulong EnterFiber(ulong fiberAddress)
+    {
+        var previous = _currentFiberAddress;
+        _currentFiberAddress = fiberAddress;
+        return previous;
+    }
+
+    public static void RestoreFiber(ulong previousFiberAddress)
+    {
+        _currentFiberAddress = previousFiberAddress;
     }
 
     public static bool RequestCurrentThreadBlock(string reason)
@@ -161,15 +185,20 @@ public static class GuestThreadExecution
         return true;
     }
 
-    public static GuestImportCallFrame EnterImportCallFrame(ulong returnRip, ulong resumeRsp)
+    public static GuestImportCallFrame EnterImportCallFrame(
+        ulong returnRip,
+        ulong resumeRsp,
+        ulong returnSlotAddress)
     {
         var previous = new GuestImportCallFrame(
             _hasCurrentImportCallFrame,
             _currentImportReturnRip,
-            _currentImportResumeRsp);
+            _currentImportResumeRsp,
+            _currentImportReturnSlotAddress);
         _hasCurrentImportCallFrame = true;
         _currentImportReturnRip = returnRip;
         _currentImportResumeRsp = resumeRsp;
+        _currentImportReturnSlotAddress = returnSlotAddress;
         return previous;
     }
 
@@ -178,6 +207,7 @@ public static class GuestThreadExecution
         _hasCurrentImportCallFrame = previous.IsValid;
         _currentImportReturnRip = previous.ReturnRip;
         _currentImportResumeRsp = previous.ResumeRsp;
+        _currentImportReturnSlotAddress = previous.ReturnSlotAddress;
     }
 
     public static bool TryGetCurrentImportCallFrame(out GuestImportCallFrame frame)
@@ -188,7 +218,11 @@ public static class GuestThreadExecution
             return false;
         }
 
-        frame = new GuestImportCallFrame(true, _currentImportReturnRip, _currentImportResumeRsp);
+        frame = new GuestImportCallFrame(
+            true,
+            _currentImportReturnRip,
+            _currentImportResumeRsp,
+            _currentImportReturnSlotAddress);
         return true;
     }
 }

src/SharpEmu.Libs/Ampr/AmprExports.cs

@@ -2,6 +2,7 @@
 // SPDX-License-Identifier: GPL-2.0-or-later
 
 using SharpEmu.HLE;
+using SharpEmu.Libs.Kernel;
 using System.Buffers.Binary;
 using System.Collections.Concurrent;
 
@@ -17,8 +18,10 @@ public static class AmprExports
     private const ulong CommandBufferAux1Offset = 0x20;
     private const ulong ReadFileRecordSize = 0x30;
     private const ulong KernelEventQueueRecordSize = 0x30;
+    private const ulong WriteAddressRecordSize = 0x20;
     private const uint ReadFileRecordType = 1;
     private const uint KernelEventQueueRecordType = 2;
+    private const uint WriteAddressRecordType = 3;
     private static readonly ConcurrentDictionary<ulong, CommandBufferState> _commandBuffers = new();
 
     private sealed class CommandBufferState
@@ -284,6 +287,18 @@ public static class AmprExports
         return (int)OrbisGen2Result.ORBIS_GEN2_OK;
     }
 
+    [SysAbiExport(
+        Nid = "C+IEj+BsAFM",
+        ExportName = "sceAmprMeasureCommandSizeWriteAddressOnCompletion",
+        Target = Generation.Gen5,
+        LibraryName = "libSceAmpr")]
+    public static int MeasureCommandSizeWriteAddressOnCompletion(CpuContext ctx)
+    {
+        TraceAmpr(ctx, "measure_write_address_complete", 0, WriteAddressRecordSize, 0);
+        ctx[CpuRegister.Rax] = WriteAddressRecordSize;
+        return (int)OrbisGen2Result.ORBIS_GEN2_OK;
+    }
+
     [SysAbiExport(
         Nid = "tZDDEo2tE5k",
         ExportName = "sceAmprCommandBufferGetSize",
@@ -362,6 +377,92 @@ public static class AmprExports
         return (int)OrbisGen2Result.ORBIS_GEN2_OK;
     }
 
+    [SysAbiExport(
+        Nid = "sJXyWHjP-F8",
+        ExportName = "sceAmprCommandBufferWriteAddressOnCompletion",
+        Target = Generation.Gen5,
+        LibraryName = "libSceAmpr")]
+    public static int CommandBufferWriteAddressOnCompletion(CpuContext ctx)
+    {
+        var commandBuffer = ctx[CpuRegister.Rdi];
+        var address = ctx[CpuRegister.Rsi];
+        var value = ctx[CpuRegister.Rdx];
+
+        if (commandBuffer == 0 || address == 0)
+        {
+            return (int)OrbisGen2Result.ORBIS_GEN2_ERROR_INVALID_ARGUMENT;
+        }
+
+        if (!AppendWriteAddressRecord(ctx, commandBuffer, address, value))
+        {
+            return (int)OrbisGen2Result.ORBIS_GEN2_ERROR_MEMORY_FAULT;
+        }
+
+        TraceAmpr(ctx, "write_address_complete", commandBuffer, address, value);
+        ctx[CpuRegister.Rax] = 0;
+        return (int)OrbisGen2Result.ORBIS_GEN2_OK;
+    }
+
+    public static int CompleteCommandBuffer(CpuContext ctx, ulong commandBuffer)
+    {
+        if (commandBuffer == 0)
+        {
+            return (int)OrbisGen2Result.ORBIS_GEN2_ERROR_INVALID_ARGUMENT;
+        }
+
+        if (!TryGetCommandBufferState(ctx, commandBuffer, out var buffer, out _, out var state) || state is null)
+        {
+            return (int)OrbisGen2Result.ORBIS_GEN2_ERROR_MEMORY_FAULT;
+        }
+
+        ulong writeOffset;
+        lock (state)
+        {
+            writeOffset = state.WriteOffset;
+        }
+
+        var offset = 0UL;
+        while (offset < writeOffset)
+        {
+            if (!TryReadUInt32(ctx, buffer + offset, out var recordType))
+            {
+                return (int)OrbisGen2Result.ORBIS_GEN2_ERROR_MEMORY_FAULT;
+            }
+
+            switch (recordType)
+            {
+                case ReadFileRecordType:
+                    offset += ReadFileRecordSize;
+                    break;
+
+                case KernelEventQueueRecordType:
+                    if (!CompleteKernelEventQueueRecord(ctx, buffer + offset))
+                    {
+                        return (int)OrbisGen2Result.ORBIS_GEN2_ERROR_MEMORY_FAULT;
+                    }
+
+                    offset += KernelEventQueueRecordSize;
+                    break;
+
+                case WriteAddressRecordType:
+                    if (!CompleteWriteAddressRecord(ctx, buffer + offset))
+                    {
+                        return (int)OrbisGen2Result.ORBIS_GEN2_ERROR_MEMORY_FAULT;
+                    }
+
+                    offset += WriteAddressRecordSize;
+                    break;
+
+                default:
+                    TraceAmpr(ctx, "complete_unknown", commandBuffer, recordType, offset);
+                    return (int)OrbisGen2Result.ORBIS_GEN2_ERROR_INVALID_ARGUMENT;
+            }
+        }
+
+        TraceAmpr(ctx, "complete", commandBuffer, buffer, writeOffset);
+        return (int)OrbisGen2Result.ORBIS_GEN2_OK;
+    }
+
     private static bool InitializeCommandBuffer(
         CpuContext ctx,
         ulong commandBuffer,
@@ -576,6 +677,17 @@ public static class AmprExports
         return AppendCommandBufferRecord(ctx, commandBuffer, record);
     }
 
+    private static bool AppendWriteAddressRecord(CpuContext ctx, ulong commandBuffer, ulong address, ulong value)
+    {
+        Span<byte> record = stackalloc byte[(int)WriteAddressRecordSize];
+        record.Clear();
+        BinaryPrimitives.WriteUInt32LittleEndian(record[0x00..], WriteAddressRecordType);
+        BinaryPrimitives.WriteUInt64LittleEndian(record[0x08..], address);
+        BinaryPrimitives.WriteUInt64LittleEndian(record[0x10..], value);
+
+        return AppendCommandBufferRecord(ctx, commandBuffer, record);
+    }
+
     private static bool AppendCommandBufferRecord(CpuContext ctx, ulong commandBuffer, ReadOnlySpan<byte> record)
     {
         if (!TryGetCommandBufferState(ctx, commandBuffer, out _, out _, out var state) || state is null)
@@ -604,6 +716,66 @@ public static class AmprExports
         return true;
     }
 
+    private static bool CompleteKernelEventQueueRecord(CpuContext ctx, ulong recordAddress)
+    {
+        Span<byte> record = stackalloc byte[(int)KernelEventQueueRecordSize];
+        if (!ctx.Memory.TryRead(recordAddress, record))
+        {
+            return false;
+        }
+
+        var filter = unchecked((short)BinaryPrimitives.ReadUInt32LittleEndian(record[0x04..]));
+        var equeue = BinaryPrimitives.ReadUInt64LittleEndian(record[0x08..]);
+        var ident = BinaryPrimitives.ReadUInt64LittleEndian(record[0x10..]);
+        var userData = BinaryPrimitives.ReadUInt64LittleEndian(record[0x18..]);
+        var data = BinaryPrimitives.ReadUInt64LittleEndian(record[0x20..]);
+        var extra = BinaryPrimitives.ReadUInt64LittleEndian(record[0x28..]);
+
+        var queuedEvent = new KernelEventQueueCompatExports.KernelQueuedEvent(
+            ident,
+            filter,
+            0x20,
+            unchecked((uint)extra),
+            data,
+            userData);
+
+        _ = KernelEventQueueCompatExports.EnqueueEvent(equeue, queuedEvent);
+        TraceAmpr(ctx, "complete_equeue", equeue, ident, data);
+        return true;
+    }
+
+    private static bool CompleteWriteAddressRecord(CpuContext ctx, ulong recordAddress)
+    {
+        Span<byte> record = stackalloc byte[(int)WriteAddressRecordSize];
+        if (!ctx.Memory.TryRead(recordAddress, record))
+        {
+            return false;
+        }
+
+        var address = BinaryPrimitives.ReadUInt64LittleEndian(record[0x08..]);
+        var value = BinaryPrimitives.ReadUInt64LittleEndian(record[0x10..]);
+        if (!ctx.TryWriteUInt64(address, value))
+        {
+            return false;
+        }
+
+        TraceAmpr(ctx, "complete_write_address", address, value, 0);
+        return true;
+    }
+
+    private static bool TryReadUInt32(CpuContext ctx, ulong address, out uint value)
+    {
+        Span<byte> buffer = stackalloc byte[sizeof(uint)];
+        if (!ctx.Memory.TryRead(address, buffer))
+        {
+            value = 0;
+            return false;
+        }
+
+        value = BinaryPrimitives.ReadUInt32LittleEndian(buffer);
+        return true;
+    }
+
     private static void TraceAmpr(CpuContext ctx, string operation, ulong commandBuffer, ulong arg0, ulong arg1)
     {
         if (!string.Equals(Environment.GetEnvironmentVariable("SHARPEMU_LOG_AMPR"), "1", StringComparison.Ordinal))

src/SharpEmu.Libs/Fiber/FiberExports.cs

@@ -230,9 +230,9 @@ public static class FiberExports
         LibraryName = "libSceFiber")]
     public static int FiberReturnToThread(CpuContext ctx)
     {
-        var fiberAddress = _currentFiberAddress;
+        var fiberAddress = ResolveCurrentFiberAddress(ctx);
         var inferredFiber = false;
-        if (fiberAddress == 0 && TryFindFiberByStack(ctx, out fiberAddress))
+        if (_currentFiberAddress == 0 && fiberAddress != 0)
         {
             inferredFiber = true;
         }
@@ -258,9 +258,9 @@ public static class FiberExports
         if (GuestThreadExecution.TryGetCurrentImportCallFrame(out var frame))
         {
             _continuations[fiberAddress] = new FiberContinuation(
-                CaptureContinuation(ctx, frame.ReturnRip, frame.ResumeRsp),
+                CaptureContinuation(ctx, frame.ReturnRip, frame.ResumeRsp, frame.ReturnSlotAddress),
                 argOnRunAddress);
-            TraceFiber($"yield{(inferredFiber ? "-inferred" : string.Empty)} fiber=0x{fiberAddress:X16} resume=0x{frame.ReturnRip:X16} rsp=0x{frame.ResumeRsp:X16} arg_out=0x{argOnRunAddress:X16}");
+            TraceFiber($"yield{(inferredFiber ? "-inferred" : string.Empty)} fiber=0x{fiberAddress:X16} resume=0x{frame.ReturnRip:X16} rsp=0x{frame.ResumeRsp:X16} return_slot=0x{frame.ReturnSlotAddress:X16} arg_out=0x{argOnRunAddress:X16}");
         }
         else
         {
@@ -284,12 +284,13 @@ public static class FiberExports
             return SetReturn(ctx, FiberErrorNull);
         }
 
-        if (_currentFiberAddress == 0)
+        var fiberAddress = ResolveCurrentFiberAddress(ctx);
+        if (fiberAddress == 0)
         {
             return SetReturn(ctx, FiberErrorPermission);
         }
 
-        return TryWriteUInt64(ctx, outAddress, _currentFiberAddress)
+        return TryWriteUInt64(ctx, outAddress, fiberAddress)
             ? SetReturn(ctx, 0)
             : SetReturn(ctx, FiberErrorInvalid);
     }
@@ -407,7 +408,7 @@ public static class FiberExports
             return SetReturn(ctx, FiberErrorNull);
         }
 
-        if (_currentFiberAddress == 0)
+        if (ResolveCurrentFiberAddress(ctx) == 0)
         {
             return SetReturn(ctx, FiberErrorPermission);
         }
@@ -537,10 +538,11 @@ public static class FiberExports
 
         if (fields.State != FiberStateIdle)
         {
+            TraceFiber($"run-state-error reason={reason} fiber=0x{fiber:X16} state=0x{fields.State:X8}");
             return SetReturn(ctx, FiberErrorState);
         }
 
-        var previousFiber = _currentFiberAddress;
+        var previousFiber = ResolveCurrentFiberAddress(ctx);
         var switchingFromFiber = isSwitch && previousFiber != 0 && previousFiber != fiber;
         if (isSwitch && previousFiber == 0)
         {
@@ -553,56 +555,111 @@ public static class FiberExports
             return SetReturn(ctx, FiberErrorPermission);
         }
 
+        var session = new FiberRunSession();
+        var ownsSession = _runSessions.TryAdd(fiber, session);
+        if (!ownsSession && !_runSessions.TryGetValue(fiber, out session))
+        {
+            return SetReturn(ctx, FiberErrorState);
+        }
+
+        FiberContinuation? previousContinuation = null;
+        GuestImportCallFrame switchFrame = default;
+        if (switchingFromFiber)
+        {
+            if (!GuestThreadExecution.TryGetCurrentImportCallFrame(out switchFrame))
+            {
+                if (ownsSession)
+                {
+                    _runSessions.TryRemove(fiber, out _);
+                }
+                TraceFiber($"switch-no-frame from=0x{previousFiber:X16} to=0x{fiber:X16} arg_out=0x{outArgumentAddress:X16}");
+                return SetReturn(ctx, FiberErrorPermission);
+            }
+
+            previousContinuation = new FiberContinuation(
+                CaptureContinuation(ctx, switchFrame.ReturnRip, switchFrame.ResumeRsp, switchFrame.ReturnSlotAddress),
+                outArgumentAddress);
+        }
+
         if (!TryWriteUInt32(ctx, fiber + FiberStateOffset, FiberStateRun))
         {
+            if (ownsSession)
+            {
+                _runSessions.TryRemove(fiber, out _);
+            }
             return SetReturn(ctx, FiberErrorInvalid);
         }
 
         if (switchingFromFiber && !TryWriteUInt32(ctx, previousFiber + FiberStateOffset, FiberStateIdle))
         {
             _ = TryWriteUInt32(ctx, fiber + FiberStateOffset, FiberStateIdle);
+            if (ownsSession)
+            {
+                _runSessions.TryRemove(fiber, out _);
+            }
             return SetReturn(ctx, FiberErrorInvalid);
         }
 
+        if (previousContinuation.HasValue)
+        {
+            _continuations[previousFiber] = previousContinuation.Value;
+            TraceFiber($"switch-save from=0x{previousFiber:X16} to=0x{fiber:X16} resume=0x{switchFrame.ReturnRip:X16} rsp=0x{switchFrame.ResumeRsp:X16} return_slot=0x{switchFrame.ReturnSlotAddress:X16} arg_out=0x{outArgumentAddress:X16}");
+        }
+
         var previousReturnRequested = _fiberReturnRequested;
         var previousReturnArgument = _fiberReturnArgument;
-        var session = new FiberRunSession();
-        _runSessions[fiber] = session;
         _currentFiberAddress = fiber;
         _fiberReturnRequested = false;
         _fiberReturnArgument = 0;
+        var previousExecutionFiber = GuestThreadExecution.EnterFiber(fiber);
+        TraceFiber($"run-enter reason={reason} fiber=0x{fiber:X16} prev=0x{previousFiber:X16} thread=0x{GuestThreadExecution.CurrentGuestThreadHandle:X16} managed={Environment.CurrentManagedThreadId} stack=0x{fields.ContextAddress:X16}");
 
         var hasContinuation = _continuations.TryGetValue(fiber, out var continuation);
-        bool callbackOk;
-        string? callbackError;
         if (hasContinuation)
         {
-            if (continuation.ArgOnRunAddress != 0 &&
+            _continuations.TryRemove(fiber, out _);
-                !TryWriteUInt64(ctx, continuation.ArgOnRunAddress, argOnRun))
+        }
+
+        TraceFiber(hasContinuation
+            ? $"run-dispatch reason={reason} fiber=0x{fiber:X16} resume=1 rip=0x{continuation.Context.Rip:X16} rsp=0x{continuation.Context.Rsp:X16} return_slot=0x{continuation.Context.ReturnSlotAddress:X16} arg_out=0x{continuation.ArgOnRunAddress:X16}"
+            : $"run-dispatch reason={reason} fiber=0x{fiber:X16} resume=0 rip=0x{fields.Entry:X16} rsp=0x{fields.ContextAddress + fields.ContextSize:X16} arg_out=0x{outArgumentAddress:X16}");
+        bool callbackOk;
+        string? callbackError;
+        try
+        {
+            if (hasContinuation)
             {
-                callbackOk = false;
+                if (continuation.ArgOnRunAddress != 0 &&
-                callbackError = $"failed to write resumed argOnRun to 0x{continuation.ArgOnRunAddress:X16}";
+                    !TryWriteUInt64(ctx, continuation.ArgOnRunAddress, argOnRun))
+                {
+                    callbackOk = false;
+                    callbackError = $"failed to write resumed argOnRun to 0x{continuation.ArgOnRunAddress:X16}";
+                }
+                else
+                {
+                    callbackOk = scheduler.TryCallGuestContinuation(
+                        ctx,
+                        continuation.Context,
+                        reason,
+                        out callbackError);
+                }
             }
             else
             {
-                callbackOk = scheduler.TryCallGuestContinuation(
+                callbackOk = scheduler.TryCallGuestFunction(
                     ctx,
-                    continuation.Context,
+                    fields.Entry,
+                    fields.ArgOnInitialize,
+                    argOnRun,
+                    fields.ContextAddress,
+                    fields.ContextSize,
                     reason,
                     out callbackError);
             }
         }
-        else
+        finally
         {
-            callbackOk = scheduler.TryCallGuestFunction(
+            GuestThreadExecution.RestoreFiber(previousExecutionFiber);
-                ctx,
-                fields.Entry,
-                fields.ArgOnInitialize,
-                argOnRun,
-                fields.ContextAddress,
-                fields.ContextSize,
-                reason,
-                out callbackError);
         }
 
         var returnRequested = _fiberReturnRequested;
@@ -613,11 +670,10 @@ public static class FiberExports
             returnArgument = sessionReturnArgument;
         }
 
-        if (!returnRequested)
+        if (ownsSession)
         {
-            _continuations.TryRemove(fiber, out _);
+            _runSessions.TryRemove(fiber, out _);
         }
-        _runSessions.TryRemove(fiber, out _);
 
         _currentFiberAddress = previousFiber;
         _fiberReturnRequested = previousReturnRequested;
@@ -644,10 +700,15 @@ public static class FiberExports
         return SetReturn(ctx, 0);
     }
 
-    private static GuestCpuContinuation CaptureContinuation(CpuContext ctx, ulong resumeRip, ulong resumeRsp) =>
+    private static GuestCpuContinuation CaptureContinuation(
+        CpuContext ctx,
+        ulong resumeRip,
+        ulong resumeRsp,
+        ulong returnSlotAddress) =>
         new(
             resumeRip,
             resumeRsp,
+            returnSlotAddress,
             ctx.Rflags == 0 ? 0x202UL : ctx.Rflags,
             ctx.FsBase,
             ctx.GsBase,
@@ -665,6 +726,24 @@ public static class FiberExports
             ctx[CpuRegister.R14],
             ctx[CpuRegister.R15]);
 
+    private static ulong ResolveCurrentFiberAddress(CpuContext ctx)
+    {
+        if (_currentFiberAddress != 0)
+        {
+            return _currentFiberAddress;
+        }
+
+        if (GuestThreadExecution.CurrentFiberAddress != 0)
+        {
+            return GuestThreadExecution.CurrentFiberAddress;
+        }
+
+        return TryFindFiberByStack(ctx, out var fiberAddress) ? fiberAddress : 0;
+    }
+
+    internal static ulong GetCurrentFiberAddressForDiagnostics(CpuContext ctx) =>
+        ResolveCurrentFiberAddress(ctx);
+
     private static int AttachContext(
         CpuContext ctx,
         ulong fiber,

src/SharpEmu.Libs/Kernel/KernelAprCompatExports.cs

@@ -2,6 +2,7 @@
 // SPDX-License-Identifier: GPL-2.0-or-later
 
 using SharpEmu.HLE;
+using SharpEmu.Libs.Ampr;
 using System.Buffers.Binary;
 using System.Collections.Concurrent;
 using System.Threading;
@@ -38,6 +39,12 @@ public static class KernelAprCompatExports
 
         _submittedCommandBuffers[submissionId] = commandBuffer;
 
+        var completionResult = AmprExports.CompleteCommandBuffer(ctx, commandBuffer);
+        if (completionResult != (int)OrbisGen2Result.ORBIS_GEN2_OK)
+        {
+            return completionResult;
+        }
+
         if (outSubmissionId != 0 && !TryWriteUInt32(ctx, outSubmissionId, submissionId))
         {
             return (int)OrbisGen2Result.ORBIS_GEN2_ERROR_MEMORY_FAULT;
@@ -93,6 +100,12 @@ public static class KernelAprCompatExports
         var submissionId = unchecked((uint)Interlocked.Increment(ref _nextSubmissionId));
         _submittedCommandBuffers[submissionId] = commandBuffer;
 
+        var completionResult = AmprExports.CompleteCommandBuffer(ctx, commandBuffer);
+        if (completionResult != (int)OrbisGen2Result.ORBIS_GEN2_OK)
+        {
+            return completionResult;
+        }
+
         TraceApr(ctx, "submit", submissionId, commandBuffer, ctx[CpuRegister.Rsi], 0);
         return (int)OrbisGen2Result.ORBIS_GEN2_OK;
     }
@@ -114,6 +127,12 @@ public static class KernelAprCompatExports
         var submissionId = unchecked((uint)Interlocked.Increment(ref _nextSubmissionId));
         _submittedCommandBuffers[submissionId] = commandBuffer;
 
+        var completionResult = AmprExports.CompleteCommandBuffer(ctx, commandBuffer);
+        if (completionResult != (int)OrbisGen2Result.ORBIS_GEN2_OK)
+        {
+            return completionResult;
+        }
+
         if (!TryWriteUInt32(ctx, outSubmissionId, submissionId))
         {
             return (int)OrbisGen2Result.ORBIS_GEN2_ERROR_MEMORY_FAULT;
@@ -154,5 +173,12 @@ public static class KernelAprCompatExports
         _ = ctx.TryReadUInt64(ctx[CpuRegister.Rsp], out returnRip);
         Console.Error.WriteLine(
             $"[LOADER][TRACE] apr.{operation}: id=0x{submissionId:X8} cmd=0x{commandBuffer:X16} priority=0x{priority:X16} aux=0x{aux:X16} ret=0x{returnRip:X16}");
+        if (aux != 0 &&
+            ctx.TryReadUInt64(aux, out var result0) &&
+            ctx.TryReadUInt64(aux + sizeof(ulong), out var result1))
+        {
+            Console.Error.WriteLine(
+                $"[LOADER][TRACE] apr.{operation}.result: addr=0x{aux:X16} q0=0x{result0:X16} q1=0x{result1:X16}");
+        }
     }
 }

src/SharpEmu.Libs/Kernel/KernelEventFlagCompatExports.cs

@@ -5,6 +5,7 @@ using System.Buffers.Binary;
 using System.Collections.Concurrent;
 using System.Text;
 using SharpEmu.HLE;
+using SharpEmu.Libs.Fiber;
 
 namespace SharpEmu.Libs.Kernel;
 
@@ -113,6 +114,7 @@ public static class KernelEventFlagCompatExports
     {
         var handle = ctx[CpuRegister.Rdi];
         var pattern = ctx[CpuRegister.Rsi];
+        var returnRip = GetCurrentReturnRip();
         if (!_eventFlags.TryGetValue(handle, out var state))
         {
             return SetReturn(ctx, OrbisGen2Result.ORBIS_GEN2_ERROR_NOT_FOUND);
@@ -122,7 +124,7 @@ public static class KernelEventFlagCompatExports
         {
             state.Bits |= pattern;
             Monitor.PulseAll(state.Gate);
-            TraceEventFlag($"set handle=0x{handle:X16} pattern=0x{pattern:X16} bits=0x{state.Bits:X16}");
+            TraceEventFlag($"set handle=0x{handle:X16} pattern=0x{pattern:X16} bits=0x{state.Bits:X16} ret=0x{returnRip:X16}");
         }
 
         return SetReturn(ctx, OrbisGen2Result.ORBIS_GEN2_OK);
@@ -203,6 +205,7 @@ public static class KernelEventFlagCompatExports
         var waitMode = unchecked((uint)ctx[CpuRegister.Rdx]);
         var resultAddress = ctx[CpuRegister.Rcx];
         var timeoutAddress = ctx[CpuRegister.R8];
+        var returnRip = GetCurrentReturnRip();
 
         if (!_eventFlags.TryGetValue(handle, out var state))
         {
@@ -232,11 +235,17 @@ public static class KernelEventFlagCompatExports
             {
                 _ = TryWriteUInt32(ctx, timeoutAddress, 0);
                 _ = TryWriteResultPattern(ctx, resultAddress, state.Bits);
-                TraceEventFlag($"wait-timeout handle=0x{handle:X16} pattern=0x{pattern:X16} timeout={timeoutUsec}");
+                TraceEventFlag($"wait-timeout handle=0x{handle:X16} pattern=0x{pattern:X16} timeout={timeoutUsec} ret=0x{returnRip:X16}");
                 return SetReturn(ctx, OrbisGen2Result.ORBIS_GEN2_ERROR_TIMED_OUT);
             }
 
-            if (!GuestThreadExecution.RequestCurrentThreadBlock("sceKernelWaitEventFlag"))
+            var currentGuestThread = GuestThreadExecution.CurrentGuestThreadHandle;
+            var currentFiber = FiberExports.GetCurrentFiberAddressForDiagnostics(ctx);
+            var managedThread = Environment.CurrentManagedThreadId;
+            var requestedBlock = GuestThreadExecution.RequestCurrentThreadBlock("sceKernelWaitEventFlag");
+            TraceEventFlag($"wait-unsatisfied handle=0x{handle:X16} pattern=0x{pattern:X16} bits=0x{state.Bits:X16} guest_thread=0x{currentGuestThread:X16} fiber=0x{currentFiber:X16} managed={managedThread} block={requestedBlock} ret=0x{returnRip:X16} frames={FormatFrameChain(ctx)}");
+            TraceEventFlag($"wait-object handle=0x{handle:X16} name='{state.Name}' {FormatGuestWaitObject(ctx)}");
+            if (!requestedBlock)
             {
                 var scheduler = GuestThreadExecution.Scheduler;
                 if (scheduler is null)
@@ -245,7 +254,7 @@ public static class KernelEventFlagCompatExports
                 }
 
                 state.WaitingThreads++;
-                TraceEventFlag($"wait-pump handle=0x{handle:X16} pattern=0x{pattern:X16} waiters={state.WaitingThreads}");
+                TraceEventFlag($"wait-pump handle=0x{handle:X16} pattern=0x{pattern:X16} waiters={state.WaitingThreads} guest_thread=0x{currentGuestThread:X16} fiber=0x{currentFiber:X16} managed={managedThread} ret=0x{returnRip:X16}");
                 var releaseWaiter = true;
                 try
                 {
@@ -265,7 +274,7 @@ public static class KernelEventFlagCompatExports
                         {
                             state.WaitingThreads = Math.Max(0, state.WaitingThreads - 1);
                             releaseWaiter = false;
-                            TraceEventFlag($"wait-wake handle=0x{handle:X16} pattern=0x{pattern:X16} bits=0x{state.Bits:X16} waiters={state.WaitingThreads}");
+                            TraceEventFlag($"wait-wake handle=0x{handle:X16} pattern=0x{pattern:X16} bits=0x{state.Bits:X16} waiters={state.WaitingThreads} ret=0x{returnRip:X16}");
                             return SetReturn(ctx, pumpedWaitResult);
                         }
 
@@ -282,7 +291,7 @@ public static class KernelEventFlagCompatExports
             }
 
             state.WaitingThreads++;
-            TraceEventFlag($"wait-block handle=0x{handle:X16} pattern=0x{pattern:X16} waiters={state.WaitingThreads}");
+            TraceEventFlag($"wait-block handle=0x{handle:X16} pattern=0x{pattern:X16} waiters={state.WaitingThreads} guest_thread=0x{currentGuestThread:X16} fiber=0x{currentFiber:X16} managed={managedThread} ret=0x{returnRip:X16}");
             return SetReturn(ctx, OrbisGen2Result.ORBIS_GEN2_OK);
         }
         finally
@@ -398,6 +407,32 @@ public static class KernelEventFlagCompatExports
         return true;
     }
 
+    private static bool TryReadUInt64(CpuContext ctx, ulong address, out ulong value)
+    {
+        Span<byte> buffer = stackalloc byte[sizeof(ulong)];
+        if (!ctx.Memory.TryRead(address, buffer))
+        {
+            value = 0;
+            return false;
+        }
+
+        value = BinaryPrimitives.ReadUInt64LittleEndian(buffer);
+        return true;
+    }
+
+    private static bool TryReadByte(CpuContext ctx, ulong address, out byte value)
+    {
+        Span<byte> buffer = stackalloc byte[1];
+        if (!ctx.Memory.TryRead(address, buffer))
+        {
+            value = 0;
+            return false;
+        }
+
+        value = buffer[0];
+        return true;
+    }
+
     private static bool TryWriteUInt32(CpuContext ctx, ulong address, uint value)
     {
         Span<byte> buffer = stackalloc byte[sizeof(uint)];
@@ -444,4 +479,101 @@ public static class KernelEventFlagCompatExports
             Console.Error.WriteLine($"[LOADER][TRACE] event_flag.{message}");
         }
     }
+
+    private static ulong GetCurrentReturnRip() =>
+        GuestThreadExecution.TryGetCurrentImportCallFrame(out var frame)
+            ? frame.ReturnRip
+            : 0UL;
+
+    private static string FormatFrameChain(CpuContext ctx)
+    {
+        Span<ulong> returns = stackalloc ulong[4];
+        var count = 0;
+        var frame = ctx[CpuRegister.Rbp];
+        for (var index = 0; index < returns.Length && frame != 0; index++)
+        {
+            if (!ctx.TryReadUInt64(frame, out var nextFrame) ||
+                !ctx.TryReadUInt64(frame + sizeof(ulong), out var returnAddress))
+            {
+                break;
+            }
+
+            returns[count++] = returnAddress;
+            if (nextFrame <= frame)
+            {
+                break;
+            }
+
+            frame = nextFrame;
+        }
+
+        return count switch
+        {
+            0 => "none",
+            1 => $"0x{returns[0]:X16}",
+            2 => $"0x{returns[0]:X16},0x{returns[1]:X16}",
+            3 => $"0x{returns[0]:X16},0x{returns[1]:X16},0x{returns[2]:X16}",
+            _ => $"0x{returns[0]:X16},0x{returns[1]:X16},0x{returns[2]:X16},0x{returns[3]:X16}",
+        };
+    }
+
+    private static string FormatGuestWaitObject(CpuContext ctx)
+    {
+        var r12 = ctx[CpuRegister.R12];
+        var r13 = ctx[CpuRegister.R13];
+        var objectAddress = r12 != 0
+            ? r12
+            : r13 >= 0xA8
+                ? r13 - 0xA8
+                : 0;
+
+        var builder = new StringBuilder(256);
+        builder.Append($"r12=0x{r12:X16} r13=0x{r13:X16}");
+        if (objectAddress == 0)
+        {
+            return builder.ToString();
+        }
+
+        builder.Append($" obj=0x{objectAddress:X16}");
+        AppendUInt32(builder, ctx, objectAddress + 0x58, "o58");
+        AppendUInt32(builder, ctx, objectAddress + 0x5C, "o5C");
+        AppendUInt64(builder, ctx, objectAddress + 0x60, "o60");
+        AppendByte(builder, ctx, objectAddress + 0x6C, "state6C");
+        AppendByte(builder, ctx, objectAddress + 0x6D, "o6D");
+        AppendByte(builder, ctx, objectAddress + 0xA0, "waitA0");
+        AppendByte(builder, ctx, objectAddress + 0xA1, "stateA1");
+        AppendByte(builder, ctx, objectAddress + 0xA2, "oA2");
+        AppendUInt64(builder, ctx, objectAddress + 0xA8, "eventA8");
+        if (r13 != 0)
+        {
+            AppendUInt64(builder, ctx, r13, "r13_0");
+            AppendUInt64(builder, ctx, r13 + 8, "r13_8");
+        }
+
+        return builder.ToString();
+    }
+
+    private static void AppendByte(StringBuilder builder, CpuContext ctx, ulong address, string name)
+    {
+        if (TryReadByte(ctx, address, out var value))
+        {
+            builder.Append($" {name}=0x{value:X2}");
+        }
+    }
+
+    private static void AppendUInt32(StringBuilder builder, CpuContext ctx, ulong address, string name)
+    {
+        if (TryReadUInt32(ctx, address, out var value))
+        {
+            builder.Append($" {name}=0x{value:X8}");
+        }
+    }
+
+    private static void AppendUInt64(StringBuilder builder, CpuContext ctx, ulong address, string name)
+    {
+        if (TryReadUInt64(ctx, address, out var value))
+        {
+            builder.Append($" {name}=0x{value:X16}");
+        }
+    }
 }

src/SharpEmu.Libs/Kernel/KernelMemoryCompatExports.cs

@@ -3,6 +3,7 @@
 
 using SharpEmu.HLE;
 using SharpEmu.Libs.Ampr;
+using System.Buffers;
 using System.Buffers.Binary;
 using System.Text;
 using System.Threading;
@@ -164,7 +165,7 @@ public static class KernelMemoryCompatExports
             var desiredAddress = AlignUp(
                 _nextVirtualAddress == 0 ? 0x1_0000_0000UL : _nextVirtualAddress,
                 effectiveAlignment);
-            if (!TryReserveGuestVirtualRange(ctx, desiredAddress, mappedLength, OrbisProtCpuReadWrite, out address) ||
+            if (!TryReserveGuestVirtualRange(ctx, desiredAddress, mappedLength, OrbisProtCpuReadWrite, effectiveAlignment, out address) ||
                 address == 0)
             {
                 return false;
@@ -597,7 +598,6 @@ public static class KernelMemoryCompatExports
             ulong NextGpArg() => vaCursor.NextGpArg();
             double NextFloatArg() => vaCursor.NextFloatArg();
             rendered = FormatString(ctx, format, NextGpArg, NextFloatArg);
-            vaCursor.Commit();
         }
 
         Console.Write(rendered);
@@ -2242,7 +2242,7 @@ public static class KernelMemoryCompatExports
             }
             else
             {
-                reserved = TryReserveGuestVirtualRange(ctx, desiredAddress, length, protection, out mappedAddress);
+                reserved = TryReserveGuestVirtualRange(ctx, desiredAddress, length, protection, effectiveAlignment, out mappedAddress);
             }
             if (ShouldTraceDirectMemory())
             {
@@ -2331,7 +2331,7 @@ public static class KernelMemoryCompatExports
             {
                 mappedAddress = requestedAddress;
             }
-            else if (!TryReserveGuestVirtualRange(ctx, desiredAddress, length, protection, out mappedAddress))
+            else if (!TryReserveGuestVirtualRange(ctx, desiredAddress, length, protection, OrbisPageSize, out mappedAddress))
             {
                 mappedAddress = requestedAddress != 0 && fixedMapping
                     ? requestedAddress
@@ -2721,7 +2721,6 @@ public static class KernelMemoryCompatExports
         ulong NextGpArg() => vaCursor.NextGpArg();
         double NextFloatArg() => vaCursor.NextFloatArg();
         var rendered = FormatString(ctx, format, NextGpArg, NextFloatArg);
-        vaCursor.Commit();
 
         var outputBytes = Encoding.UTF8.GetBytes(rendered);
         return WriteSnprintfOutput(ctx, destination, bufferSize, outputBytes);
@@ -2768,7 +2767,6 @@ public static class KernelMemoryCompatExports
             ulong NextGpArg() => vaCursor.NextGpArg();
             double NextFloatArg() => vaCursor.NextFloatArg();
             rendered = FormatString(ctx, format, NextGpArg, NextFloatArg);
-            vaCursor.Commit();
         }
 
         TraceWidePrintf(ctx, "vswprintf", destination, bufferSize, format, rendered);
@@ -3552,13 +3550,6 @@ public static class KernelMemoryCompatExports
                 : 0.0;
         }
 
-        public void Commit()
-        {
-            _ = TryWriteUInt32Compat(_ctx, _vaListAddress + 0, _gpOffset);
-            _ = TryWriteUInt32Compat(_ctx, _vaListAddress + 4, _fpOffset);
-            _ = TryWriteUInt64Compat(_ctx, _vaListAddress + 8, _overflowArgArea);
-        }
-
         public uint GpOffset => _gpOffset;
 
         public uint FpOffset => _fpOffset;
@@ -3613,6 +3604,7 @@ public static class KernelMemoryCompatExports
         ulong desiredAddress,
         ulong length,
         int protection,
+        ulong alignment,
         out ulong mappedAddress)
     {
         mappedAddress = 0;
@@ -3625,40 +3617,51 @@ public static class KernelMemoryCompatExports
         {
             object memoryObject = ctx.Memory;
             MethodInfo? allocateAt = null;
+            MethodInfo? allocateAtOrAbove = null;
             var allocateAtHasAllowAlternativeArg = false;
             for (var depth = 0; depth < 4; depth++)
             {
                 foreach (var candidate in memoryObject.GetType().GetMethods(BindingFlags.Public | BindingFlags.Instance))
                 {
-                    if (!string.Equals(candidate.Name, "AllocateAt", StringComparison.Ordinal))
-                    {
-                        continue;
-                    }
-
                     var parameters = candidate.GetParameters();
-                    if (parameters.Length == 3 &&
+                    if (string.Equals(candidate.Name, "TryAllocateAtOrAbove", StringComparison.Ordinal) &&
-                        parameters[0].ParameterType == typeof(ulong) &&
+                        parameters.Length == 5 &&
-                        parameters[1].ParameterType == typeof(ulong) &&
-                        parameters[2].ParameterType == typeof(bool))
-                    {
-                        allocateAt = candidate;
-                        allocateAtHasAllowAlternativeArg = false;
-                        break;
-                    }
-
-                    if (parameters.Length == 4 &&
                         parameters[0].ParameterType == typeof(ulong) &&
                         parameters[1].ParameterType == typeof(ulong) &&
                         parameters[2].ParameterType == typeof(bool) &&
-                        parameters[3].ParameterType == typeof(bool))
+                        parameters[3].ParameterType == typeof(ulong) &&
+                        parameters[4].ParameterType == typeof(ulong).MakeByRefType())
+                    {
+                        allocateAtOrAbove = candidate;
+                    }
+                    else if (string.Equals(candidate.Name, "AllocateAt", StringComparison.Ordinal))
+                    {
+                        if (parameters.Length == 3 &&
+                            parameters[0].ParameterType == typeof(ulong) &&
+                            parameters[1].ParameterType == typeof(ulong) &&
+                            parameters[2].ParameterType == typeof(bool))
+                        {
+                            allocateAt = candidate;
+                            allocateAtHasAllowAlternativeArg = false;
+                        }
+                        else if (parameters.Length == 4 &&
+                            parameters[0].ParameterType == typeof(ulong) &&
+                            parameters[1].ParameterType == typeof(ulong) &&
+                            parameters[2].ParameterType == typeof(bool) &&
+                            parameters[3].ParameterType == typeof(bool))
+                        {
+                            allocateAt = candidate;
+                            allocateAtHasAllowAlternativeArg = true;
+                        }
+                    }
+
+                    if (allocateAtOrAbove is not null && allocateAt is not null)
                     {
-                        allocateAt = candidate;
-                        allocateAtHasAllowAlternativeArg = true;
                         break;
                     }
                 }
 
-                if (allocateAt is not null)
+                if (allocateAtOrAbove is not null || allocateAt is not null)
                 {
                     break;
                 }
@@ -3678,15 +3681,27 @@ public static class KernelMemoryCompatExports
                 memoryObject = innerValue;
             }
 
+            var executable = (protection & OrbisProtCpuExec) != 0;
+            if (allocateAtOrAbove is not null)
+            {
+                var searchArgs = new object[] { desiredAddress, length, executable, alignment, 0UL };
+                var searchResult = allocateAtOrAbove.Invoke(memoryObject, searchArgs);
+                if (searchResult is bool trueValue && trueValue &&
+                    searchArgs[4] is ulong searchedAddress && searchedAddress != 0)
+                {
+                    mappedAddress = searchedAddress;
+                    return true;
+                }
+            }
+
             if (allocateAt is null)
             {
                 Console.Error.WriteLine($"[LOADER][TRACE] reserve range: AllocateAt missing on {ctx.Memory.GetType().FullName}");
                 return false;
             }
 
-            var executable = (protection & OrbisProtCpuExec) != 0;
             var invokeArgs = allocateAtHasAllowAlternativeArg
-                ? new object[] { desiredAddress, length, executable, true }
+                ? new object[] { desiredAddress, length, executable, false }
                 : new object[] { desiredAddress, length, executable };
             var result = allocateAt.Invoke(memoryObject, invokeArgs);
             if (result is not ulong allocated || allocated == 0)
@@ -3959,6 +3974,16 @@ public static class KernelMemoryCompatExports
     }
 
     private static bool TryReadCString(CpuContext ctx, ulong address, ulong maxLength, out byte[] bytes)
+    {
+        return TryReadBytesUntilNull(ctx, address, maxLength, 1_048_576, out bytes);
+    }
+
+    private static bool TryReadBytesUntilNull(
+        CpuContext ctx,
+        ulong address,
+        ulong maxLength,
+        int hardLimit,
+        out byte[] bytes)
     {
         bytes = Array.Empty<byte>();
         if (address == 0)
@@ -3966,26 +3991,65 @@ public static class KernelMemoryCompatExports
             return false;
         }
 
-        var limit = (int)Math.Min(maxLength, 1_048_576UL);
+        var limit = (int)Math.Min(maxLength, (ulong)Math.Max(0, hardLimit));
-        var buffer = new List<byte>(Math.Min(limit, 256));
+        if (limit == 0)
-        Span<byte> one = stackalloc byte[1];
-        for (var i = 0; i < limit; i++)
         {
-            if (!TryReadCompat(ctx, address + (ulong)i, one))
+            return true;
+        }
+
+        const int maxChunkSize = 4096;
+        var chunk = GC.AllocateUninitializedArray<byte>(Math.Min(maxChunkSize, limit));
+        var writer = new ArrayBufferWriter<byte>(Math.Min(limit, 256));
+        ulong offset = 0;
+        while (offset < (ulong)limit)
+        {
+            var current = address + offset;
+            if (current < address)
+            {
+                return false;
+            }
+
+            var pageRemaining = maxChunkSize - (int)(current & (maxChunkSize - 1));
+            var remaining = (int)Math.Min((ulong)limit - offset, (ulong)Math.Min(chunk.Length, pageRemaining));
+            var span = chunk.AsSpan(0, remaining);
+            if (TryReadCompat(ctx, current, span))
+            {
+                var nulIndex = span.IndexOf((byte)0);
+                var copyLength = nulIndex >= 0 ? nulIndex : remaining;
+                if (copyLength > 0)
+                {
+                    span[..copyLength].CopyTo(writer.GetSpan(copyLength));
+                    writer.Advance(copyLength);
+                }
+
+                if (nulIndex >= 0)
+                {
+                    bytes = writer.WrittenSpan.ToArray();
+                    return true;
+                }
+
+                offset += (ulong)remaining;
+                continue;
+            }
+
+            Span<byte> one = stackalloc byte[1];
+            if (!TryReadCompat(ctx, current, one))
             {
                 return false;
             }
 
             if (one[0] == 0)
             {
-                bytes = buffer.ToArray();
+                bytes = writer.WrittenSpan.ToArray();
                 return true;
             }
 
-            buffer.Add(one[0]);
+            one.CopyTo(writer.GetSpan(1));
+            writer.Advance(1);
+            offset++;
         }
 
-        bytes = buffer.ToArray();
+        bytes = writer.WrittenSpan.ToArray();
         return true;
     }
 
@@ -4165,25 +4229,12 @@ public static class KernelMemoryCompatExports
             return false;
         }
 
-        var buffer = new List<byte>(Math.Min(maxLength, 256));
+        if (!TryReadBytesUntilNull(ctx, address, (ulong)maxLength, maxLength, out var bytes))
-        Span<byte> one = stackalloc byte[1];
-        for (var i = 0; i < maxLength; i++)
         {
-            if (!TryReadCompat(ctx, address + (ulong)i, one))
+            return false;
-            {
-                return false;
-            }
-
-            if (one[0] == 0)
-            {
-                value = Encoding.UTF8.GetString(buffer.ToArray());
-                return true;
-            }
-
-            buffer.Add(one[0]);
         }
 
-        value = Encoding.UTF8.GetString(buffer.ToArray());
+        value = Encoding.UTF8.GetString(bytes);
         return true;
     }
 

src/SharpEmu.Libs/Kernel/KernelRuntimeCompatExports.cs

@@ -2,6 +2,7 @@
 // SPDX-License-Identifier: GPL-2.0-or-later
 
 using SharpEmu.HLE;
+using SharpEmu.Libs.Fiber;
 using System.Buffers.Binary;
 using System.Collections.Generic;
 using System.Diagnostics;
@@ -32,6 +33,8 @@ public static class KernelRuntimeCompatExports
     private const ulong DefaultKernelTscFrequency = 10_000_000UL;
     private const ulong PrtAreaStartAddress = 0x0000001000000000UL;
     private const ulong PrtAreaSize = 0x000000EC00000000UL;
+    private const int MapFlagFixed = 0x10;
+    private const ulong DefaultVirtualRangeAlignment = 0x4000UL;
     private const int AioInitParamSize = 0x3C;
     private const uint MemCommit = 0x1000;
     private const uint MemReserve = 0x2000;
@@ -73,6 +76,8 @@ public static class KernelRuntimeCompatExports
             return (int)OrbisGen2Result.ORBIS_GEN2_OK;
         }
 
+        GuestThreadExecution.Scheduler?.Pump(ctx, "sceKernelUsleep");
+
         if (micros < 1000)
         {
             Thread.Yield();
@@ -108,8 +113,14 @@ public static class KernelRuntimeCompatExports
             lockText = $"0x{lockValue:X16}";
         }
 
+        var returnRip = GuestThreadExecution.TryGetCurrentImportCallFrame(out var frame)
+            ? frame.ReturnRip
+            : 0UL;
+        var thread = GuestThreadExecution.CurrentGuestThreadHandle;
+        var fiber = FiberExports.GetCurrentFiberAddressForDiagnostics(ctx);
+
         Console.Error.WriteLine(
-            $"[LOADER][TRACE] usleep#{count}: usec={micros} rbx=0x{rbx:X16} lock@+F78=0x{lockAddress:X16}:{lockText} r13=0x{ctx[CpuRegister.R13]:X16} r14=0x{ctx[CpuRegister.R14]:X16} r15=0x{ctx[CpuRegister.R15]:X16}");
+            $"[LOADER][TRACE] usleep#{count}: usec={micros} ret=0x{returnRip:X16} thread=0x{thread:X16} fiber=0x{fiber:X16} rbx=0x{rbx:X16} lock@+F78=0x{lockAddress:X16}:{lockText} r13=0x{ctx[CpuRegister.R13]:X16} r14=0x{ctx[CpuRegister.R14]:X16} r15=0x{ctx[CpuRegister.R15]:X16}");
     }
 
     [SysAbiExport(
@@ -653,7 +664,7 @@ public static class KernelRuntimeCompatExports
     {
         var inOutAddressPointer = ctx[CpuRegister.Rdi];
         var length = ctx[CpuRegister.Rsi];
-        var _flags = ctx[CpuRegister.Rdx];
+        var flags = unchecked((int)ctx[CpuRegister.Rdx]);
         var alignment = ctx[CpuRegister.Rcx];
         if (inOutAddressPointer == 0 || length == 0)
         {
@@ -665,7 +676,8 @@ public static class KernelRuntimeCompatExports
             return (int)OrbisGen2Result.ORBIS_GEN2_ERROR_MEMORY_FAULT;
         }
 
-        var effectiveAlignment = alignment == 0 ? 0x10000UL : alignment;
+        var effectiveAlignment = alignment == 0 ? DefaultVirtualRangeAlignment : alignment;
+        var fixedMapping = (flags & MapFlagFixed) != 0;
         ulong desiredAddress;
         lock (_stateGate)
         {
@@ -674,13 +686,13 @@ public static class KernelRuntimeCompatExports
                 : AlignUp(_nextReservedVirtualBase, effectiveAlignment);
         }
 
-        if (!TryReserveVirtualRange(ctx, desiredAddress, length, out var mappedAddress))
+        if (!TryReserveVirtualRange(ctx, desiredAddress, length, effectiveAlignment, allowSearch: !fixedMapping, out var mappedAddress))
         {
             return (int)OrbisGen2Result.ORBIS_GEN2_ERROR_NOT_FOUND;
         }
 
         Console.Error.WriteLine(
-            $"[LOADER][TRACE] reserve_virtual_range: req=0x{requestedAddress:X16} desired=0x{desiredAddress:X16} mapped=0x{mappedAddress:X16} len=0x{length:X16}");
+            $"[LOADER][TRACE] reserve_virtual_range: req=0x{requestedAddress:X16} desired=0x{desiredAddress:X16} mapped=0x{mappedAddress:X16} len=0x{length:X16} flags=0x{flags:X8} align=0x{effectiveAlignment:X16}");
 
         if (!ctx.TryWriteUInt64(inOutAddressPointer, mappedAddress))
         {
@@ -964,14 +976,12 @@ public static class KernelRuntimeCompatExports
     public static int StackCheckFail(CpuContext ctx)
     {
         var count = Interlocked.Increment(ref _stackChkFailCount);
-        if (count <= 8)
+        Console.Error.WriteLine(
-        {
+            $"[LOADER][ERROR] __stack_chk_fail#{count}: rip=0x{ctx.Rip:X16} rdi=0x{ctx[CpuRegister.Rdi]:X16}");
-            Console.Error.WriteLine(
+        var result = (int)OrbisGen2Result.ORBIS_GEN2_ERROR_CPU_TRAP;
-                $"[LOADER][WARNING] __stack_chk_fail recovery#{count}: rip=0x{ctx.Rip:X16} rdi=0x{ctx[CpuRegister.Rdi]:X16}");
+        GuestThreadExecution.RequestCurrentEntryExit("__stack_chk_fail", result);
-        }
+        ctx[CpuRegister.Rax] = unchecked((ulong)result);
-
+        return result;
-        ctx[CpuRegister.Rax] = 0;
-        return (int)OrbisGen2Result.ORBIS_GEN2_OK;
     }
 
     [SysAbiExport(
@@ -1656,7 +1666,13 @@ public static class KernelRuntimeCompatExports
     [DllImport("kernel32.dll", SetLastError = true)]
     private static extern nint VirtualAlloc(nint lpAddress, nuint dwSize, uint flAllocationType, uint flProtect);
 
-    private static bool TryReserveVirtualRange(CpuContext ctx, ulong desiredAddress, ulong length, out ulong mappedAddress)
+    private static bool TryReserveVirtualRange(
+        CpuContext ctx,
+        ulong desiredAddress,
+        ulong length,
+        ulong alignment,
+        bool allowSearch,
+        out ulong mappedAddress)
     {
         mappedAddress = 0;
         if (length == 0)
@@ -1668,40 +1684,51 @@ public static class KernelRuntimeCompatExports
         {
             object memoryObject = ctx.Memory;
             MethodInfo? allocateAt = null;
+            MethodInfo? allocateAtOrAbove = null;
             var allocateAtHasAllowAlternativeArg = false;
             for (var depth = 0; depth < 4; depth++)
             {
                 foreach (var candidate in memoryObject.GetType().GetMethods(BindingFlags.Public | BindingFlags.Instance))
                 {
-                    if (!string.Equals(candidate.Name, "AllocateAt", StringComparison.Ordinal))
-                    {
-                        continue;
-                    }
-
                     var parameters = candidate.GetParameters();
-                    if (parameters.Length == 3 &&
+                    if (string.Equals(candidate.Name, "TryAllocateAtOrAbove", StringComparison.Ordinal) &&
-                        parameters[0].ParameterType == typeof(ulong) &&
+                        parameters.Length == 5 &&
-                        parameters[1].ParameterType == typeof(ulong) &&
-                        parameters[2].ParameterType == typeof(bool))
-                    {
-                        allocateAt = candidate;
-                        allocateAtHasAllowAlternativeArg = false;
-                        break;
-                    }
-
-                    if (parameters.Length == 4 &&
                         parameters[0].ParameterType == typeof(ulong) &&
                         parameters[1].ParameterType == typeof(ulong) &&
                         parameters[2].ParameterType == typeof(bool) &&
-                        parameters[3].ParameterType == typeof(bool))
+                        parameters[3].ParameterType == typeof(ulong) &&
+                        parameters[4].ParameterType == typeof(ulong).MakeByRefType())
+                    {
+                        allocateAtOrAbove = candidate;
+                    }
+                    else if (string.Equals(candidate.Name, "AllocateAt", StringComparison.Ordinal))
+                    {
+                        if (parameters.Length == 3 &&
+                            parameters[0].ParameterType == typeof(ulong) &&
+                            parameters[1].ParameterType == typeof(ulong) &&
+                            parameters[2].ParameterType == typeof(bool))
+                        {
+                            allocateAt = candidate;
+                            allocateAtHasAllowAlternativeArg = false;
+                        }
+                        else if (parameters.Length == 4 &&
+                            parameters[0].ParameterType == typeof(ulong) &&
+                            parameters[1].ParameterType == typeof(ulong) &&
+                            parameters[2].ParameterType == typeof(bool) &&
+                            parameters[3].ParameterType == typeof(bool))
+                        {
+                            allocateAt = candidate;
+                            allocateAtHasAllowAlternativeArg = true;
+                        }
+                    }
+
+                    if (allocateAtOrAbove is not null && allocateAt is not null)
                     {
-                        allocateAt = candidate;
-                        allocateAtHasAllowAlternativeArg = true;
                         break;
                     }
                 }
 
-                if (allocateAt is not null)
+                if (allocateAtOrAbove is not null || allocateAt is not null)
                 {
                     break;
                 }
@@ -1721,6 +1748,18 @@ public static class KernelRuntimeCompatExports
                 memoryObject = innerValue;
             }
 
+            if (allowSearch && allocateAtOrAbove is not null)
+            {
+                var searchArgs = new object[] { desiredAddress, length, false, alignment, 0UL };
+                var searchResult = allocateAtOrAbove.Invoke(memoryObject, searchArgs);
+                if (searchResult is bool trueValue && trueValue &&
+                    searchArgs[4] is ulong searchedAddress && searchedAddress != 0)
+                {
+                    mappedAddress = searchedAddress;
+                    return true;
+                }
+            }
+
             if (allocateAt is null)
             {
                 Console.Error.WriteLine($"[LOADER][TRACE] reserve_virtual_range: AllocateAt missing on {ctx.Memory.GetType().FullName}");
@@ -1728,7 +1767,7 @@ public static class KernelRuntimeCompatExports
             }
 
             var invokeArgs = allocateAtHasAllowAlternativeArg
-                ? new object[] { desiredAddress, length, false, true }
+                ? new object[] { desiredAddress, length, false, false }
                 : new object[] { desiredAddress, length, false };
             var result = allocateAt.Invoke(memoryObject, invokeArgs);
             if (result is not ulong allocated || allocated == 0)