diff --git a/build.sh b/build.sh
new file mode 100755
index 00000000..f0be8730
--- /dev/null
+++ b/build.sh
@@ -0,0 +1,4 @@
+#!/bin/bash
+mkdir build && cd build
+cmake .. -DCMAKE_BUILD_TYPE=release -DCMAKE_C_COMPILER=/usr/bin/gcc -DCMAKE_CXX_COMPILER=/usr/bin/g++
+cd ..
diff --git a/src/CMakeLists.txt b/src/CMakeLists.txt
index 005cfa2a..3505d6e0 100644
--- a/src/CMakeLists.txt
+++ b/src/CMakeLists.txt
@@ -21,8 +21,9 @@ if(MSVC)
     # _SILENCE_ALL_CXX17_DEPRECATION_WARNINGS
 elseif(UNIX)
     add_definitions(-fms-extensions)
-    add_definitions(-fms-compatibility-version=19.14)
-    add_definitions(-fdelayed-template-parsing)
+    # add_definitions(-fms-compatibility-version=19.14)
+    # add_definitions(-fdelayed-template-parsing)
+    add_definitions(-fpermissive)
     add_definitions(-DVK_USE_PLATFORM_XLIB_KHR) # legacy. Do we need to support XLIB surfaces?
     add_definitions(-DVK_USE_PLATFORM_XCB_KHR)
     add_definitions(-maes)
diff --git a/src/Cafe/CMakeLists.txt b/src/Cafe/CMakeLists.txt
index ff0f3e11..1e3fe306 100644
--- a/src/Cafe/CMakeLists.txt
+++ b/src/Cafe/CMakeLists.txt
@@ -1,10 +1,16 @@
 project(CemuCafe)
 
 include_directories(".")
+
+if((CMAKE_C_COMPILER_ID MATCHES "GNU") OR (CMAKE_C_COMPILER_ID MATCHES "Clang"))
+	add_compile_options(-mssse3 -mavx2)
+endif()
+
 file(GLOB_RECURSE CPP_FILES *.cpp)
 file(GLOB_RECURSE H_FILES *.h)
 add_library(CemuCafe ${CPP_FILES} ${H_FILES})
 
+
 set_property(TARGET CemuCafe PROPERTY MSVC_RUNTIME_LIBRARY "MultiThreaded$<$<CONFIG:Debug>:Debug>")
 
 target_precompile_headers(CemuCafe PRIVATE ../Common/precompiled.h)
@@ -23,7 +29,6 @@ target_link_libraries(CemuCafe glslang SPIRV)
 target_link_libraries(CemuCafe ih264d zarchive)
 #target_link_libraries(CemuCafe zstd::libzstd_static)
 
-
 IF(WIN32)
 target_link_libraries(CemuCafe iphlpapi)
 ENDIF()
\ No newline at end of file
diff --git a/src/Cafe/CafeSystem.cpp b/src/Cafe/CafeSystem.cpp
index 53a9ec6a..a07942d2 100644
--- a/src/Cafe/CafeSystem.cpp
+++ b/src/Cafe/CafeSystem.cpp
@@ -282,7 +282,7 @@ struct
 
 static_assert(sizeof(SharedDataEntry) == 0x1C);
 
-__declspec(dllexport) uint32 loadSharedData()
+DLLEXPORT uint32 loadSharedData()
 {
 	// check if font files are dumped
 	bool hasAllShareddataFiles = true;
diff --git a/src/Cafe/GameProfile/GameProfile.cpp b/src/Cafe/GameProfile/GameProfile.cpp
index 8e1186c9..8d3f739e 100644
--- a/src/Cafe/GameProfile/GameProfile.cpp
+++ b/src/Cafe/GameProfile/GameProfile.cpp
@@ -22,7 +22,7 @@ struct gameProfileBooleanOption_t
 * If the option exists, true is returned.
 * The boolean is stored in *optionValue
 */
-__declspec(dllexport) bool gameProfile_loadBooleanOption(IniParser* iniParser, char* optionName, gameProfileBooleanOption_t* option)
+DLLEXPORT bool gameProfile_loadBooleanOption(IniParser* iniParser, char* optionName, gameProfileBooleanOption_t* option)
 {
 	auto option_value = iniParser->FindOption(optionName);
 	option->isPresent = false;
@@ -81,7 +81,7 @@ bool gameProfile_loadBooleanOption2(IniParser& iniParser, const char* optionName
 * Attempts to load a integer option
 * Allows to specify min and max value (error is logged if out of range and default value is picked)
 */
-__declspec(dllexport) bool gameProfile_loadIntegerOption(IniParser* iniParser, const char* optionName, gameProfileIntegerOption_t* option, sint32 defaultValue, sint32 minVal, sint32 maxVal)
+DLLEXPORT bool gameProfile_loadIntegerOption(IniParser* iniParser, const char* optionName, gameProfileIntegerOption_t* option, sint32 defaultValue, sint32 minVal, sint32 maxVal)
 {
 	auto option_value = iniParser->FindOption(optionName);
 	option->isPresent = false;
@@ -167,7 +167,7 @@ bool gameProfile_loadEnumOption(IniParser& iniParser, const char* optionName, st
 }
 
 #pragma optimize( "", off )  
-__declspec(dllexport) __declspec(noinline) void gameProfile_categoryBegin(IniParser* iniParser)
+DLLEXPORT NOINLINE void gameProfile_categoryBegin(IniParser* iniParser)
 {
 	// do nothing
 }
@@ -381,11 +381,11 @@ void GameProfile::Reset()
 }
 
 // legacy code for Cemuhook
-__declspec(dllexport) char* gameProfile_loadStringOption(IniParser* iniParser, char* optionName)
+DLLEXPORT char* gameProfile_loadStringOption(IniParser* iniParser, char* optionName)
 {
 	return nullptr;
 }
-__declspec(dllexport) char* gameProfile_getCurrentCategoryName(IniParser* iniParser)
+DLLEXPORT char* gameProfile_getCurrentCategoryName(IniParser* iniParser)
 {
 	return nullptr;
 }
@@ -396,7 +396,7 @@ struct gpNamedOptionEntry_t
 	sint32 value;
 };
 
-__declspec(dllexport) bool gameProfile_loadIntegerNamedOption(IniParser* iniParser, char* optionName, gameProfileIntegerOption_t* option, sint32 defaultValue, const gpNamedOptionEntry_t* nameValues, sint32 numNameValues)
+DLLEXPORT bool gameProfile_loadIntegerNamedOption(IniParser* iniParser, char* optionName, gameProfileIntegerOption_t* option, sint32 defaultValue, const gpNamedOptionEntry_t* nameValues, sint32 numNameValues)
 {
 	return false;
 }
\ No newline at end of file
diff --git a/src/Cafe/GameProfile/GameProfile.h b/src/Cafe/GameProfile/GameProfile.h
index 00b4e0d1..10e90fea 100644
--- a/src/Cafe/GameProfile/GameProfile.h
+++ b/src/Cafe/GameProfile/GameProfile.h
@@ -65,4 +65,4 @@ private:
 };
 extern std::unique_ptr<GameProfile> g_current_game_profile;
 
-__declspec(dllexport) void gameProfile_load();
+DLLEXPORT void gameProfile_load();
diff --git a/src/Cafe/GraphicPack/GraphicPack.cpp b/src/Cafe/GraphicPack/GraphicPack.cpp
index b55ac82a..bdf21e4b 100644
--- a/src/Cafe/GraphicPack/GraphicPack.cpp
+++ b/src/Cafe/GraphicPack/GraphicPack.cpp
@@ -10,7 +10,7 @@ typedef struct
 }graphicPack_t;
 
 // scans the graphic pack directory for shaders
-__declspec(dllexport) void graphicPack_loadGraphicPackShaders(graphicPack_t* gp, wchar_t* graphicPackPath)
+DLLEXPORT void graphicPack_loadGraphicPackShaders(graphicPack_t* gp, wchar_t* graphicPackPath)
 {
 	// this function is part of the deprecated/removed v1 graphic pack code
 	// as of Cemuhook 0.5.7.3 this function must exist with a minimum length for detour
@@ -21,7 +21,7 @@ __declspec(dllexport) void graphicPack_loadGraphicPackShaders(graphicPack_t* gp,
 }
 
 // for cemuhook compatibility only
-__declspec(dllexport) bool config_isGraphicPackEnabled(uint64 id)
+DLLEXPORT bool config_isGraphicPackEnabled(uint64 id)
 {
 	forceLog_printf("STUB4");
 	forceLog_printf("STUB5");
diff --git a/src/Cafe/GraphicPack/GraphicPack2.cpp b/src/Cafe/GraphicPack/GraphicPack2.cpp
index 62765f2c..0a04b154 100644
--- a/src/Cafe/GraphicPack/GraphicPack2.cpp
+++ b/src/Cafe/GraphicPack/GraphicPack2.cpp
@@ -601,13 +601,13 @@ void GraphicPack2::LoadShaders()
 
 #pragma optimize( "", off )
 
-DLLEXPORT __declspec(noinline) void GraphicPack2_notifyActivate(GraphicPack2* gp, ExpressionParser* ep)
+DLLEXPORT NOINLINE void GraphicPack2_notifyActivate(GraphicPack2* gp, ExpressionParser* ep)
 {
 	// for Cemuhook
 	int placeholder = 0xDEADDEAD;
 }
 
-DLLEXPORT __declspec(noinline) void GraphicPack2_notifyDeactivate(GraphicPack2* gp)
+DLLEXPORT NOINLINE void GraphicPack2_notifyDeactivate(GraphicPack2* gp)
 {
 	// for Cemuhook
 	int placeholder = 0xDEADDEAD;
diff --git a/src/Cafe/GraphicPack/GraphicPack2.h b/src/Cafe/GraphicPack/GraphicPack2.h
index b02dbf84..f038fb64 100644
--- a/src/Cafe/GraphicPack/GraphicPack2.h
+++ b/src/Cafe/GraphicPack/GraphicPack2.h
@@ -166,8 +166,8 @@ public:
 	static bool DeactivateGraphicPack(const std::shared_ptr<GraphicPack2>& graphic_pack);
 	static void ClearGraphicPacks();
 private:
-	__declspec(dllexport) bool Activate();
-	__declspec(dllexport) bool Deactivate();
+	DLLEXPORT bool Activate();
+	DLLEXPORT bool Deactivate();
 
 	static std::vector<std::shared_ptr<GraphicPack2>> s_graphic_packs;
 	static std::vector<std::shared_ptr<GraphicPack2>> s_active_graphic_packs;
diff --git a/src/Cafe/HW/Espresso/Interpreter/PPCInterpreterMain.cpp b/src/Cafe/HW/Espresso/Interpreter/PPCInterpreterMain.cpp
index c130ec8f..7cd27ab5 100644
--- a/src/Cafe/HW/Espresso/Interpreter/PPCInterpreterMain.cpp
+++ b/src/Cafe/HW/Espresso/Interpreter/PPCInterpreterMain.cpp
@@ -34,7 +34,7 @@ PPCInterpreter_t* PPCInterpreter_getCurrentInstance()
 	return ppcInterpreterCurrentInstance;
 }
 
-__declspec(noinline) uint64 PPCInterpreter_getMainCoreCycleCounter()
+NOINLINE uint64 PPCInterpreter_getMainCoreCycleCounter()
 {
 	return PPCTimer_getFromRDTSC();
 }
diff --git a/src/Cafe/HW/Espresso/PPCScheduler.cpp b/src/Cafe/HW/Espresso/PPCScheduler.cpp
index ebe6a473..e64b61a2 100644
--- a/src/Cafe/HW/Espresso/PPCScheduler.cpp
+++ b/src/Cafe/HW/Espresso/PPCScheduler.cpp
@@ -100,7 +100,7 @@ PPCInterpreter_t* PPCCore_executeCallbackInternal(uint32 functionMPTR)
 	return hCPU;
 }
 
-__declspec(dllexport) void PPCCore_executeCallback(uint32 functionMPTR)
+DLLEXPORT void PPCCore_executeCallback(uint32 functionMPTR)
 {
 	PPCCore_executeCallbackInternal(functionMPTR);
 }
diff --git a/src/Cafe/HW/Espresso/PPCTimer.cpp b/src/Cafe/HW/Espresso/PPCTimer.cpp
index 160b776a..24339b32 100644
--- a/src/Cafe/HW/Espresso/PPCTimer.cpp
+++ b/src/Cafe/HW/Espresso/PPCTimer.cpp
@@ -7,9 +7,11 @@
 
 #if BOOST_OS_LINUX > 0
 static __inline__
-unsigned __int64 _umul128(unsigned __int64,
-                          unsigned __int64,
-                          unsigned __int64*);
+uint64 _umul128(uint64 multiplier, uint64 multiplicand, uint64 *highProduct) {
+	unsigned __int128 x = (unsigned __int128)multiplier * (unsigned __int128)multiplicand;
+	*highProduct = (x >> 64);
+	return x & 0xFFFFFFFFFFFFFFFF;
+}
 #endif
 
 uint64 _rdtscLastMeasure = 0;
@@ -49,7 +51,7 @@ uint64 PPCTimer_estimateRDTSCFrequency()
 		forceLog_printf("Invariant TSC not supported");
 
 	_mm_mfence();
-	unsigned __int64 tscStart = __rdtsc();
+	uint64 tscStart = __rdtsc();
 	unsigned int startTime = GetTickCount();
 	HRTick startTick = HighResolutionTimer::now().getTick();
 	// wait roughly 3 seconds
@@ -61,7 +63,7 @@ uint64 PPCTimer_estimateRDTSCFrequency()
 	}
 	_mm_mfence();
 	HRTick stopTick = HighResolutionTimer::now().getTick();
-	unsigned __int64 tscEnd = __rdtsc();
+	uint64 tscEnd = __rdtsc();
 	// derive frequency approximation from measured time difference
 	uint64 tsc_diff = tscEnd - tscStart;
 	uint64 hrtFreq = 0;
diff --git a/src/Cafe/HW/Espresso/Recompiler/PPCRecompiler.cpp b/src/Cafe/HW/Espresso/Recompiler/PPCRecompiler.cpp
index 75adee35..6b9bb542 100644
--- a/src/Cafe/HW/Espresso/Recompiler/PPCRecompiler.cpp
+++ b/src/Cafe/HW/Espresso/Recompiler/PPCRecompiler.cpp
@@ -580,7 +580,7 @@ void PPCRecompiler_init()
 	__cpuid(cpuInfo, 0x1);
 	hasMOVBESupport = ((cpuInfo[2] >> 22) & 1) != 0;
 	hasAVXSupport = ((cpuInfo[2] >> 28) & 1) != 0;
-	__cpuidex(cpuInfo, 0x7, 0);
+	__cpuidex1(cpuInfo, 0x7, 0);
 	hasBMI2Support = ((cpuInfo[1] >> 8) & 1) != 0;
 
 	forceLog_printf("Recompiler initialized. CPU extensions: %s%s%s", hasLZCNTSupport ? "LZCNT " : "", hasMOVBESupport ? "MOVBE " : "", hasAVXSupport ? "AVX " : "");
diff --git a/src/Cafe/HW/Espresso/Recompiler/PPCRecompiler.h b/src/Cafe/HW/Espresso/Recompiler/PPCRecompiler.h
index f87778de..a87d53a4 100644
--- a/src/Cafe/HW/Espresso/Recompiler/PPCRecompiler.h
+++ b/src/Cafe/HW/Espresso/Recompiler/PPCRecompiler.h
@@ -345,20 +345,20 @@ typedef struct
 	PPCRecFunction_t* ppcRecompilerFuncTable[PPC_REC_ALIGN_TO_4MB(PPC_REC_CODE_AREA_SIZE/4)]; // one virtual-function pointer for each potential ppc instruction
 	PPCREC_JUMP_ENTRY ppcRecompilerDirectJumpTable[PPC_REC_ALIGN_TO_4MB(PPC_REC_CODE_AREA_SIZE/4)]; // lookup table for ppc offset to native code function
 	// x64 data
-	uint64 __declspec(align(16)) _x64XMM_xorNegateMaskBottom[2];
-	uint64 __declspec(align(16)) _x64XMM_xorNegateMaskPair[2];
-	uint64 __declspec(align(16)) _x64XMM_xorNOTMask[2];
-	uint64 __declspec(align(16)) _x64XMM_andAbsMaskBottom[2];
-	uint64 __declspec(align(16)) _x64XMM_andAbsMaskPair[2];
-	uint32 __declspec(align(16)) _x64XMM_andFloatAbsMaskBottom[4];
-	uint64 __declspec(align(16)) _x64XMM_singleWordMask[2];
-	double __declspec(align(16)) _x64XMM_constDouble1_1[2];
-	double __declspec(align(16)) _x64XMM_constDouble0_0[2];
-	float  __declspec(align(16)) _x64XMM_constFloat0_0[2];
-	float  __declspec(align(16)) _x64XMM_constFloat1_1[2];
-	float  __declspec(align(16)) _x64XMM_constFloatMin[2];
-	uint32 __declspec(align(16)) _x64XMM_flushDenormalMask1[4];
-	uint32 __declspec(align(16)) _x64XMM_flushDenormalMaskResetSignBits[4];
+	uint64 ALIGN(16) _x64XMM_xorNegateMaskBottom[2];
+	uint64 ALIGN(16) _x64XMM_xorNegateMaskPair[2];
+	uint64 ALIGN(16) _x64XMM_xorNOTMask[2];
+	uint64 ALIGN(16) _x64XMM_andAbsMaskBottom[2];
+	uint64 ALIGN(16) _x64XMM_andAbsMaskPair[2];
+	uint32 ALIGN(16) _x64XMM_andFloatAbsMaskBottom[4];
+	uint64 ALIGN(16) _x64XMM_singleWordMask[2];
+	double ALIGN(16) _x64XMM_constDouble1_1[2];
+	double ALIGN(16) _x64XMM_constDouble0_0[2];
+	float  ALIGN(16) _x64XMM_constFloat0_0[2];
+	float  ALIGN(16) _x64XMM_constFloat1_1[2];
+	float  ALIGN(16) _x64XMM_constFloatMin[2];
+	uint32 ALIGN(16) _x64XMM_flushDenormalMask1[4];
+	uint32 ALIGN(16) _x64XMM_flushDenormalMaskResetSignBits[4];
 	// PSQ load/store scale tables
 	double _psq_ld_scale_ps0_ps1[64 * 2];
 	double _psq_ld_scale_ps0_1[64 * 2];
@@ -369,10 +369,10 @@ typedef struct
 	uint32 _x64XMM_mxCsr_ftzOff;
 }PPCRecompilerInstanceData_t;
 
-extern __declspec(dllexport) PPCRecompilerInstanceData_t* ppcRecompilerInstanceData;
+extern DLLEXPORT PPCRecompilerInstanceData_t* ppcRecompilerInstanceData;
 extern bool ppcRecompilerEnabled;
 
-__declspec(dllexport) void PPCRecompiler_init();
+DLLEXPORT void PPCRecompiler_init();
 
 void PPCRecompiler_allocateRange(uint32 startAddress, uint32 size);
 
@@ -385,10 +385,10 @@ extern void ATTR_MS_ABI (*PPCRecompiler_leaveRecompilerCode_unvisited)();
 #define PPC_REC_INVALID_FUNCTION	((PPCRecFunction_t*)-1)
 
 // CPUID
-extern __declspec(dllexport) bool hasLZCNTSupport;
-extern __declspec(dllexport) bool hasMOVBESupport;
-extern __declspec(dllexport) bool hasBMI2Support;
-extern __declspec(dllexport) bool hasAVXSupport;
+extern DLLEXPORT bool hasLZCNTSupport;
+extern DLLEXPORT bool hasMOVBESupport;
+extern DLLEXPORT bool hasBMI2Support;
+extern DLLEXPORT bool hasAVXSupport;
 
 // todo - move some of the stuff above into PPCRecompilerInternal.h
 
diff --git a/src/Cafe/HW/Espresso/Recompiler/PPCRecompilerX64.cpp b/src/Cafe/HW/Espresso/Recompiler/PPCRecompilerX64.cpp
index cda0e60a..7b36a775 100644
--- a/src/Cafe/HW/Espresso/Recompiler/PPCRecompilerX64.cpp
+++ b/src/Cafe/HW/Espresso/Recompiler/PPCRecompilerX64.cpp
@@ -2221,7 +2221,10 @@ void PPCRecompilerX64Gen_imlInstruction_r_name(PPCRecFunction_t* PPCRecFunction,
 		else if (sprIndex == SPR_XER)
 			x64Emit_mov_reg64_mem32(x64GenContext, tempToRealRegister(imlInstruction->op_r_name.registerIndex), REG_RSP, offsetof(PPCInterpreter_t, spr.XER));
 		else if (sprIndex >= SPR_UGQR0 && sprIndex <= SPR_UGQR7)
-			x64Emit_mov_reg64_mem32(x64GenContext, tempToRealRegister(imlInstruction->op_r_name.registerIndex), REG_RSP, offsetof(PPCInterpreter_t, spr.UGQR[sprIndex - SPR_UGQR0]));
+		{
+			sint32 memOffset = offsetof(PPCInterpreter_t, spr.UGQR) + sizeof(PPCInterpreter_t::spr.UGQR[0]) * (sprIndex - SPR_UGQR0);
+			x64Emit_mov_reg64_mem32(x64GenContext, tempToRealRegister(imlInstruction->op_r_name.registerIndex), REG_RSP, memOffset);
+		}
 		else
 			assert_dbg();
 		//x64Emit_mov_reg64_mem32(x64GenContext, tempToRealRegister(imlInstruction->op_r_name.registerIndex), REG_RSP, offsetof(PPCInterpreter_t, spr)+sizeof(uint32)*(name-PPCREC_NAME_SPR0));
@@ -2239,7 +2242,7 @@ void PPCRecompilerX64Gen_imlInstruction_name_r(PPCRecFunction_t* PPCRecFunction,
 	}
 	else if( name >= PPCREC_NAME_SPR0 && name < PPCREC_NAME_SPR0+999 )
 	{
-		uint32 sprIndex = (name - PPCREC_NAME_SPR0);
+		const uint32 sprIndex = (name - PPCREC_NAME_SPR0);
 		if (sprIndex == SPR_LR)
 			x64Emit_mov_mem32_reg64(x64GenContext, REG_RSP, offsetof(PPCInterpreter_t, spr.LR), tempToRealRegister(imlInstruction->op_r_name.registerIndex));
 		else if (sprIndex == SPR_CTR)
@@ -2247,7 +2250,10 @@ void PPCRecompilerX64Gen_imlInstruction_name_r(PPCRecFunction_t* PPCRecFunction,
 		else if (sprIndex == SPR_XER)
 			x64Emit_mov_mem32_reg64(x64GenContext, REG_RSP, offsetof(PPCInterpreter_t, spr.XER), tempToRealRegister(imlInstruction->op_r_name.registerIndex));
 		else if (sprIndex >= SPR_UGQR0 && sprIndex <= SPR_UGQR7)
-			x64Emit_mov_mem32_reg64(x64GenContext, REG_RSP, offsetof(PPCInterpreter_t, spr.UGQR[sprIndex-SPR_UGQR0]), tempToRealRegister(imlInstruction->op_r_name.registerIndex));
+		{
+			sint32 memOffset = offsetof(PPCInterpreter_t, spr.UGQR) + sizeof(PPCInterpreter_t::spr.UGQR[0]) * (sprIndex - SPR_UGQR0);
+			x64Emit_mov_mem32_reg64(x64GenContext, REG_RSP, memOffset, tempToRealRegister(imlInstruction->op_r_name.registerIndex));
+		}
 		else
 			assert_dbg();	
 	}
diff --git a/src/Cafe/HW/Latte/Core/LatteOverlay.cpp b/src/Cafe/HW/Latte/Core/LatteOverlay.cpp
index ae483c32..102520db 100644
--- a/src/Cafe/HW/Latte/Core/LatteOverlay.cpp
+++ b/src/Cafe/HW/Latte/Core/LatteOverlay.cpp
@@ -516,7 +516,7 @@ void LatteOverlay_translateScreenPosition(ScreenPosition pos, const Vector2f& wi
 		direction = -1;
 		break;
 	default:
-		__assume(false);
+		ASSUME(false);
 	}
 }
 
diff --git a/src/Cafe/HW/Latte/Core/LatteTextureCache.cpp b/src/Cafe/HW/Latte/Core/LatteTextureCache.cpp
index ae7ac4d0..8ec54c05 100644
--- a/src/Cafe/HW/Latte/Core/LatteTextureCache.cpp
+++ b/src/Cafe/HW/Latte/Core/LatteTextureCache.cpp
@@ -430,7 +430,7 @@ void LatteTC_UnloadAllTextures()
 /*
  * Asynchronous way to invalidate textures
  */
-__declspec(dllexport) void gpu7Texture_forceInvalidateByImagePtr(MPTR imagePtr)
+DLLEXPORT void gpu7Texture_forceInvalidateByImagePtr(MPTR imagePtr)
 {
 	// deprecated. Texture cache heuristics are now good enough to detect moving frames
 }
diff --git a/src/Cafe/HW/Latte/LegacyShaderDecompiler/LatteDecompilerInternal.h b/src/Cafe/HW/Latte/LegacyShaderDecompiler/LatteDecompilerInternal.h
index 4c2c2efd..a35da277 100644
--- a/src/Cafe/HW/Latte/LegacyShaderDecompiler/LatteDecompilerInternal.h
+++ b/src/Cafe/HW/Latte/LegacyShaderDecompiler/LatteDecompilerInternal.h
@@ -124,7 +124,7 @@ struct LatteDecompilerCFInstruction
 	LatteDecompilerCFInstruction(LatteDecompilerCFInstruction&& mE) = default;
 #else
 	LatteDecompilerCFInstruction(const LatteDecompilerCFInstruction& mE) = default;
-	LatteDecompilerCFInstruction(const LatteDecompilerCFInstruction&& mE) = default;
+	constexpr LatteDecompilerCFInstruction(LatteDecompilerCFInstruction&& mE) = default;
 #endif
 	LatteDecompilerCFInstruction& operator=(LatteDecompilerCFInstruction&& mE) = default;
 };
diff --git a/src/Cafe/HW/Latte/Renderer/OpenGL/OpenGLRenderer.cpp b/src/Cafe/HW/Latte/Renderer/OpenGL/OpenGLRenderer.cpp
index d89cbcac..548e4923 100644
--- a/src/Cafe/HW/Latte/Renderer/OpenGL/OpenGLRenderer.cpp
+++ b/src/Cafe/HW/Latte/Renderer/OpenGL/OpenGLRenderer.cpp
@@ -1458,7 +1458,7 @@ void OpenGLRenderer::shader_bind(RendererShader* shader)
 		prevGeometryShaderProgram = program;
 		break;
 	default:
-		__assume(false);
+		ASSUME(false);
 	}
 
 	catchOpenGLError();
@@ -1489,7 +1489,7 @@ void OpenGLRenderer::shader_unbind(RendererShader::ShaderType shaderType)
 		prevGeometryShaderProgram = -1;
 		break;
 	default: 
-		__assume(false);
+		ASSUME(false);
 	}
 }
 
diff --git a/src/Cafe/HW/Latte/Renderer/OpenGL/RendererShaderGL.cpp b/src/Cafe/HW/Latte/Renderer/OpenGL/RendererShaderGL.cpp
index 60dd99ea..8ea170ca 100644
--- a/src/Cafe/HW/Latte/Renderer/OpenGL/RendererShaderGL.cpp
+++ b/src/Cafe/HW/Latte/Renderer/OpenGL/RendererShaderGL.cpp
@@ -270,7 +270,7 @@ void RendererShaderGL::ShaderCacheLoading_begin(uint64 cacheTitleId)
 		usePrecompiled = false;
 		break;
 	default:
-		__assume(false);
+		ASSUME(false);
 	}
 
 	forceLog_printf("Using precompiled shaders: %s", usePrecompiled ? "true" : "false");
diff --git a/src/Cafe/HW/Latte/Renderer/OpenGL/RendererShaderGL.h b/src/Cafe/HW/Latte/Renderer/OpenGL/RendererShaderGL.h
index 5c929998..1c0753dc 100644
--- a/src/Cafe/HW/Latte/Renderer/OpenGL/RendererShaderGL.h
+++ b/src/Cafe/HW/Latte/Renderer/OpenGL/RendererShaderGL.h
@@ -1,7 +1,7 @@
 #pragma once
 
 #include "Cafe/HW/Latte/Renderer/RendererShader.h"
-#include "Common\GLInclude\GLInclude.h"
+#include "Common/GLInclude/GLInclude.h"
 
 class RendererShaderGL : public RendererShader
 {
diff --git a/src/Cafe/HW/Latte/Renderer/Vulkan/RendererShaderVk.cpp b/src/Cafe/HW/Latte/Renderer/Vulkan/RendererShaderVk.cpp
index b107e56d..2dfd2009 100644
--- a/src/Cafe/HW/Latte/Renderer/Vulkan/RendererShaderVk.cpp
+++ b/src/Cafe/HW/Latte/Renderer/Vulkan/RendererShaderVk.cpp
@@ -7,7 +7,7 @@
 #endif
 
 #include <glslang/Public/ShaderLang.h>
-#if GLSLANG_VERSION_LESS_OR_EQUAL_TO(11, 0, 0)
+#if 1 //GLSLANG_VERSION_LESS_OR_EQUAL_TO(11, 0, 0)
 #include <glslang/SPIRV/GlslangToSpv.h>
 #else
 #include <SPIRV/Logger.h>
@@ -131,7 +131,7 @@ const TBuiltInResource DefaultTBuiltInResource = {
 	/* .maxDualSourceDrawBuffersEXT = */ 1,
 #endif
 
-	/* .limits = */ {
+	/* .limits = */
 	/* .nonInductiveForLoops = */ 1,
 	/* .whileLoops = */ 1,
 	/* .doWhileLoops = */ 1,
@@ -141,7 +141,6 @@ const TBuiltInResource DefaultTBuiltInResource = {
 	/* .generalSamplerIndexing = */ 1,
 	/* .generalVariableIndexing = */ 1,
 	/* .generalConstantMatrixVectorIndexing = */ 1,
-}
 };
 
 class _ShaderVkThreadPool
@@ -470,4 +469,4 @@ void RendererShaderVk::ShaderCacheLoading_end()
 {
 	// keep g_spirvCache open since we will write to it while the game is running
 	s_isLoadingShadersVk = false;
-}
\ No newline at end of file
+}
diff --git a/src/Cafe/HW/Latte/Renderer/Vulkan/VulkanRenderer.cpp b/src/Cafe/HW/Latte/Renderer/Vulkan/VulkanRenderer.cpp
index a792fc69..56e8888c 100644
--- a/src/Cafe/HW/Latte/Renderer/Vulkan/VulkanRenderer.cpp
+++ b/src/Cafe/HW/Latte/Renderer/Vulkan/VulkanRenderer.cpp
@@ -3409,7 +3409,7 @@ VkDescriptorSetInfo::~VkDescriptorSetInfo()
 		break;
 	}
 	default:
-		__assume(false);
+		ASSUME(false);
 	}
 	// update global stats
 	performanceMonitor.vk.numDescriptorSamplerTextures.decrement(statsNumSamplerTextures);
diff --git a/src/Cafe/HW/Latte/Renderer/Vulkan/VulkanRendererCore.cpp b/src/Cafe/HW/Latte/Renderer/Vulkan/VulkanRendererCore.cpp
index e47b5bf0..cefec598 100644
--- a/src/Cafe/HW/Latte/Renderer/Vulkan/VulkanRendererCore.cpp
+++ b/src/Cafe/HW/Latte/Renderer/Vulkan/VulkanRendererCore.cpp
@@ -528,7 +528,7 @@ uint64 VulkanRenderer::GetDescriptorSetStateHash(LatteDecompilerShader* shader)
 			texUnitRegIndex += Latte::REGADDR::SQ_TEX_RESOURCE_WORD0_N_GS;
 			break;
 		default:
-			__assume(false);
+			ASSUME(false);
 		}
 
 		auto texture = m_state.boundTexture[hostTextureUnit];
@@ -590,7 +590,7 @@ VkDescriptorSetInfo* VulkanRenderer::draw_getOrCreateDescriptorSet(PipelineInfo*
 		break;
 	}
 	default:
-		__assume(false);
+		ASSUME(false);
 	}
 
 	// create new descriptor set
@@ -646,7 +646,7 @@ VkDescriptorSetInfo* VulkanRenderer::draw_getOrCreateDescriptorSet(PipelineInfo*
 			texUnitRegIndex += Latte::REGADDR::SQ_TEX_RESOURCE_WORD0_N_GS;
 			break;
 		default:
-			__assume(false);
+			ASSUME(false);
 		}
 
 		auto textureView = m_state.boundTexture[hostTextureUnit];
@@ -996,7 +996,7 @@ VkDescriptorSetInfo* VulkanRenderer::draw_getOrCreateDescriptorSet(PipelineInfo*
 		break;
 	}
 	default:
-		__assume(false);
+		ASSUME(false);
 	}
 
 	return dsInfo;
diff --git a/src/Cafe/HW/MMU/MMU.cpp b/src/Cafe/HW/MMU/MMU.cpp
index 3982d8d8..1d746d68 100644
--- a/src/Cafe/HW/MMU/MMU.cpp
+++ b/src/Cafe/HW/MMU/MMU.cpp
@@ -389,7 +389,7 @@ uint8 memory_readU8(uint32 address)
 	return *(uint8*)(memory_getPointerFromVirtualOffset(address));
 }
 
-__declspec(dllexport) void* memory_getBase()
+DLLEXPORT void* memory_getBase()
 {
 	return memory_base;
 }
diff --git a/src/Cafe/OS/RPL/rpl.cpp b/src/Cafe/OS/RPL/rpl.cpp
index 07cf3d0d..cf17a42c 100644
--- a/src/Cafe/OS/RPL/rpl.cpp
+++ b/src/Cafe/OS/RPL/rpl.cpp
@@ -46,8 +46,8 @@ ChunkedFlatAllocator<64 * 1024> g_heapTrampolineArea;
 
 std::vector<rplDependency_t*> rplDependencyList = std::vector<rplDependency_t*>();
 
-__declspec(dllexport) RPLModule* rplModuleList[256];
-__declspec(dllexport) sint32 rplModuleCount = 0;
+DLLEXPORT RPLModule* rplModuleList[256];
+DLLEXPORT sint32 rplModuleCount = 0;
 
 uint32 _currentTLSModuleIndex = 1; // value 0 is reserved
 
diff --git a/src/Cafe/OS/RPL/rpl.h b/src/Cafe/OS/RPL/rpl.h
index d7eae552..8be6d9af 100644
--- a/src/Cafe/OS/RPL/rpl.h
+++ b/src/Cafe/OS/RPL/rpl.h
@@ -14,7 +14,7 @@ MPTR RPLLoader_AllocateCodeSpace(uint32 size, uint32 alignment);
 uint32 RPLLoader_GetMaxCodeOffset();
 uint32 RPLLoader_GetDataAllocatorAddr();
 
-__declspec(dllexport) RPLModule* rpl_loadFromMem(uint8* rplData, sint32 size, char* name);
+DLLEXPORT RPLModule* rpl_loadFromMem(uint8* rplData, sint32 size, char* name);
 uint32 rpl_mapHLEImport(RPLModule* rplLoaderContext, const char* rplName, const char* funcName, bool functionMustExist);
 void RPLLoader_Link();
 
@@ -52,4 +52,4 @@ uint32 RPLLoader_MakePPCCallable(void(*ppcCallableExport)(struct PPCInterpreter_
 
 // elf loader
 
-uint32 ELF_LoadFromMemory(uint8* elfData, sint32 size, const char* name);
\ No newline at end of file
+uint32 ELF_LoadFromMemory(uint8* elfData, sint32 size, const char* name);
diff --git a/src/Cafe/OS/common/OSCommon.cpp b/src/Cafe/OS/common/OSCommon.cpp
index dbc9d16e..77f1fa6d 100644
--- a/src/Cafe/OS/common/OSCommon.cpp
+++ b/src/Cafe/OS/common/OSCommon.cpp
@@ -99,7 +99,7 @@ void osLib_addFunctionInternal(const char* libraryName, const char* functionName
 	s_osFunctionTable->emplace_back(libHashA, libHashB, funcHashA, funcHashB, fmt::format("{}.{}", libraryName, functionName), PPCInterpreter_registerHLECall(osFunction));
 }
 
-__declspec(dllexport) void osLib_registerHLEFunction(const char* libraryName, const char* functionName, void(*osFunction)(PPCInterpreter_t* hCPU))
+DLLEXPORT void osLib_registerHLEFunction(const char* libraryName, const char* functionName, void(*osFunction)(PPCInterpreter_t* hCPU))
 {
 	osLib_addFunctionInternal(libraryName, functionName, osFunction);
 }
diff --git a/src/Cafe/OS/libs/coreinit/coreinit_MEM_BlockHeap.cpp b/src/Cafe/OS/libs/coreinit/coreinit_MEM_BlockHeap.cpp
index 1738677e..e0bcbcaf 100644
--- a/src/Cafe/OS/libs/coreinit/coreinit_MEM_BlockHeap.cpp
+++ b/src/Cafe/OS/libs/coreinit/coreinit_MEM_BlockHeap.cpp
@@ -398,7 +398,7 @@ namespace coreinit
 		else
 			track = (MEMBlockHeapTrackDEPR*)memory_getPointerFromVirtualOffsetAllowNull(_swapEndianU32(blockHeapHead->headBlock));
 
-		cemu_assert_debug(__popcnt(alignment) == 1); // not a supported alignment value
+		cemu_assert_debug(POPCNT(alignment) == 1); // not a supported alignment value
 		while (track)
 		{
 			if (track->isFree != _swapEndianU32(0))
diff --git a/src/Cafe/OS/libs/coreinit/coreinit_Thread.cpp b/src/Cafe/OS/libs/coreinit/coreinit_Thread.cpp
index 43eae921..a57b4473 100644
--- a/src/Cafe/OS/libs/coreinit/coreinit_Thread.cpp
+++ b/src/Cafe/OS/libs/coreinit/coreinit_Thread.cpp
@@ -45,8 +45,8 @@ namespace coreinit
 
 	bool g_isMulticoreMode;
 
-	__declspec(thread) uint32 t_assignedCoreIndex;
-	__declspec(thread) Fiber* t_schedulerFiber;
+	THREAD_LOCAL uint32 t_assignedCoreIndex;
+	THREAD_LOCAL Fiber* t_schedulerFiber;
 
 	struct OSHostThread
 	{
diff --git a/src/Cafe/OS/libs/coreinit/coreinit_Time.cpp b/src/Cafe/OS/libs/coreinit/coreinit_Time.cpp
index 823d4938..fd51f81e 100644
--- a/src/Cafe/OS/libs/coreinit/coreinit_Time.cpp
+++ b/src/Cafe/OS/libs/coreinit/coreinit_Time.cpp
@@ -28,7 +28,7 @@ namespace coreinit
 		osLib_returnFromFunction64(hCPU, osTime);
 	}
 
-	__declspec(noinline) uint64 coreinit_getTimeBase_dummy()
+	NOINLINE uint64 coreinit_getTimeBase_dummy()
 	{
 		return __rdtsc();
 	}
diff --git a/src/Cafe/OS/libs/nlibcurl/nlibcurl.cpp b/src/Cafe/OS/libs/nlibcurl/nlibcurl.cpp
index 4208c1ff..5e9f8b51 100644
--- a/src/Cafe/OS/libs/nlibcurl/nlibcurl.cpp
+++ b/src/Cafe/OS/libs/nlibcurl/nlibcurl.cpp
@@ -201,8 +201,8 @@ static_assert(sizeof(CURLMsg_t) <= 0xC, "sizeof(CURLMsg_t)");
 
 size_t header_callback(char* buffer, size_t size, size_t nitems, void* userdata);
 
-__declspec(thread) PPCConcurrentQueue<QueueMsg_t>* g_callerQueue;
-__declspec(thread) ConcurrentQueue<QueueMsg_t>* g_threadQueue;
+THREAD_LOCAL PPCConcurrentQueue<QueueMsg_t>* g_callerQueue;
+THREAD_LOCAL ConcurrentQueue<QueueMsg_t>* g_threadQueue;
 void CurlWorkerThread(CURL_t* curl, PPCConcurrentQueue<QueueMsg_t>* callerQueue, ConcurrentQueue<QueueMsg_t>* threadQueue)
 {
 	g_callerQueue = callerQueue;
diff --git a/src/Cafe/TitleList/MetaInfo.cpp b/src/Cafe/TitleList/MetaInfo.cpp
index f3e362b0..8fda420b 100644
--- a/src/Cafe/TitleList/MetaInfo.cpp
+++ b/src/Cafe/TitleList/MetaInfo.cpp
@@ -174,7 +174,7 @@ std::unique_ptr<uint8[]> MetaInfo::GetIcon(uint32& size) const
 		return nullptr;
 	}
 	default:
-		__assume(false);
+		ASSUME(false);
 	}
 	return nullptr;
 }
diff --git a/src/Cemu/Logging/CemuLogging.cpp b/src/Cemu/Logging/CemuLogging.cpp
index c22ad30e..fb9b8a57 100644
--- a/src/Cemu/Logging/CemuLogging.cpp
+++ b/src/Cemu/Logging/CemuLogging.cpp
@@ -243,7 +243,7 @@ void cafeLog_logW(uint32 type, const wchar_t* format, ...)
 		LoggingWindow::Log(it->second, logTempStr);
 }
 
-__declspec(dllexport) void cemuLog_log()
+DLLEXPORT void cemuLog_log()
 {
 	typedef void(*VoidFunc)();
 	const VoidFunc func = (VoidFunc)cafeLog_log;
diff --git a/src/Common/SysAllocator.h b/src/Common/SysAllocator.h
index 1807fb24..6c77e780 100644
--- a/src/Common/SysAllocator.h
+++ b/src/Common/SysAllocator.h
@@ -1,6 +1,7 @@
 #pragma once
 
 #include <vector>
+
 uint32 coreinit_allocFromSysArea(uint32 size, uint32 alignment);
 class SysAllocatorBase;
 
@@ -195,4 +196,4 @@ private:
 
 	MEMPTR<T> m_sysMem;
 	T m_tempData;
-};
\ No newline at end of file
+};
diff --git a/src/Common/linux/platform.cpp b/src/Common/linux/platform.cpp
index 592a665f..86973c7f 100644
--- a/src/Common/linux/platform.cpp
+++ b/src/Common/linux/platform.cpp
@@ -16,8 +16,8 @@ void (__cpuid)(int __cpuVal[4], unsigned int __leaf)
 }
 #undef __cpuid
 
-void __cpuidex (int __cpuid_info[4], int __leaf, int __subleaf)
+void __cpuidex1(int __cpuid_info[4], int __leaf, int __subleaf)
 {
   __cpuid_count (__leaf, __subleaf, __cpuid_info[0], __cpuid_info[1],
 		 __cpuid_info[2], __cpuid_info[3]);
-}
\ No newline at end of file
+}
diff --git a/src/Common/linux/platform.h b/src/Common/linux/platform.h
index f8aeff56..735954a5 100644
--- a/src/Common/linux/platform.h
+++ b/src/Common/linux/platform.h
@@ -45,7 +45,7 @@ inline uint32_t GetExceptionError()
 
 // cpu id (somewhat hacky, reorganize later)
 void (__cpuid)(int __cpuVal[4], unsigned int __leaf);
-void __cpuidex (int __cpuid_info[4], int __leaf, int __subleaf);
+void __cpuidex1 (int __cpuid_info[4], int __leaf, int __subleaf);
 
 // placeholder
 uint32_t GetTickCount();
diff --git a/src/Common/precompiled.h b/src/Common/precompiled.h
index 0ec725a6..e9b56bf6 100644
--- a/src/Common/precompiled.h
+++ b/src/Common/precompiled.h
@@ -166,6 +166,16 @@ inline sint16 _swapEndianS16(sint16 v)
     return (sint16)(((uint16)v >> 8) | ((uint16)v << 8));
 }
 
+inline uint64 _rotl64 (uint64 x, sint8 r)
+{
+    return (x << r) | (x >> (64 - r));
+}
+
+inline uint64 _rotr64 (uint64 x, sint8 r)
+{
+    return (x >> r) | (x << (64 - r));
+}
+
 typedef uint8_t BYTE;
 typedef uint32_t DWORD;
 typedef int32_t LONG;
@@ -200,11 +210,26 @@ typedef union _LARGE_INTEGER {
 
 // macros
 #if BOOST_OS_WINDOWS
-#define DLLEXPORT __declspec(dllexport)
+    #define DLLEXPORT __declspec(dllexport)
+    #define DLLIMPORT  __declspec(dllimport)
+    #define DEBUG_BREAK __debugbreak()
+    #define ALIGN(N) __declspec(align(N))
+    #define NOINLINE __declspec(noinline)
+    #define ASSUME(X) __assume((X)
+    #define THREAD_LOCAL __declspec(thread)
+    #define POPCNT(X) __popcnt((X))
 #else
-#define DLLEXPORT __declspec(dllexport)
+    #define DLLEXPORT
+    #define DLLIMPORT
+    #define DEBUG_BREAK
+    #define ALIGN(N) __attribute__((aligned (N)))
+    #define NOINLINE __attribute__((noinline))
+    #define ASSUME(X)
+    #define THREAD_LOCAL __thread
+    #define POPCNT(X) __builtin_popcount((X))
 #endif
 
+
 template <typename T1, typename T2>
 constexpr bool HAS_FLAG(T1 flags, T2 test_flag) { return (flags & (T1)test_flag) == (T1)test_flag; }
 template <typename T1, typename T2>
@@ -286,7 +311,7 @@ inline void cemu_assert(bool _condition)
 {
     if ((_condition) == false)
     {
-        __debugbreak();
+        DEBUG_BREAK;
     }
 }
 
@@ -307,32 +332,32 @@ inline void cemu_assert_suspicious()
 
 inline void cemu_assert_error()
 {
-	__debugbreak();
+	DEBUG_BREAK;
 }
 #else
 inline void cemu_assert_debug(bool _condition)
 {
     if ((_condition) == false)
-        __debugbreak();
+        DEBUG_BREAK;
 }
 
 inline void cemu_assert_unimplemented()
 {
-    __debugbreak();
+    DEBUG_BREAK;
 }
 
 inline void cemu_assert_suspicious()
 {
-    __debugbreak();
+    DEBUG_BREAK;
 }
 
 inline void cemu_assert_error()
 {
-    __debugbreak();
+    DEBUG_BREAK;
 }
 #endif
 
-#define assert_dbg() __debugbreak() // old style unconditional generic assert
+#define assert_dbg() DEBUG_BREAK // old style unconditional generic assert
 
 // Some string conversion helpers because C++20 std::u8string is too cumbersome to use in practice
 // mixing string types generally causes loads of issues and many of the libraries we use dont expose interfaces for u8string
@@ -383,7 +408,7 @@ public:
 template<typename T>
 bool future_is_ready(std::future<T>& f)
 {
-#ifdef __clang__
+#if defined(__clang__) || defined(__GNUC__)
     return f.wait_for(std::chrono::nanoseconds(0)) == std::future_status::ready;
 #else
 	return f._Is_ready();
@@ -399,7 +424,7 @@ std::atomic<T>* _rawPtrToAtomic(T* ptr)
     return reinterpret_cast<std::atomic<T>*>(ptr);
 }
 
-#if __clang__
+#if defined(__clang__) || defined(__GNUC__)
 #define ATTR_MS_ABI __attribute__((ms_abi))
 #else
 #define ATTR_MS_ABI
@@ -420,7 +445,7 @@ inline uint32 GetTitleIdLow(uint64 titleId)
 	return titleId & 0xFFFFFFFF;
 }
 
-#ifdef __clang__
+#if defined(__clang__) || defined(__GNUC__)
 #define memcpy_dwords(__dest, __src, __numDwords) memcpy((__dest), (__src), (__numDwords) * sizeof(uint32))
 #define memcpy_qwords(__dest, __src, __numQwords) memcpy((__dest), (__src), (__numQwords) * sizeof(uint64))
 #else
diff --git a/src/cemuhook/wxCemuhookExports.cpp b/src/cemuhook/wxCemuhookExports.cpp
index d39e22dc..adcd8a0d 100644
--- a/src/cemuhook/wxCemuhookExports.cpp
+++ b/src/cemuhook/wxCemuhookExports.cpp
@@ -31,28 +31,28 @@
 #define CHECK_FOR_WX_EVT_STRING(strVar, strConst) if (strcmp(strVar, #strConst) == 0){ return static_cast<int>(strConst); }
 
 
-__declspec(dllexport) wxEvtHandler* wxEvtHandler_Initialize(uint8_t* allocMemory)
+DLLEXPORT wxEvtHandler* wxEvtHandler_Initialize(uint8_t* allocMemory)
 {
 	wxEvtHandler* handler = new (allocMemory) wxEvtHandler();
 	return handler;
 }
 
-__declspec(dllexport) void wxEvtHandler_Connect(wxEvtHandler* eventSource, int id, int lastId, int eventType, wxObjectEventFunction func, wxObject* userData, wxEvtHandler* eventSink)
+DLLEXPORT void wxEvtHandler_Connect(wxEvtHandler* eventSource, int id, int lastId, int eventType, wxObjectEventFunction func, wxObject* userData, wxEvtHandler* eventSink)
 {
 	eventSource->Connect(id, lastId, eventType, func, userData, eventSink);
 }
 
-__declspec(dllexport) void wxEvtHandler_Disconnect(wxEvtHandler* eventSource, int id, int lastId, int eventType, wxObjectEventFunction func, wxObject* userData, wxEvtHandler* eventSink)
+DLLEXPORT void wxEvtHandler_Disconnect(wxEvtHandler* eventSource, int id, int lastId, int eventType, wxObjectEventFunction func, wxObject* userData, wxEvtHandler* eventSink)
 {
 	eventSource->Disconnect(id, lastId, eventType, func, userData, eventSink);
 }
 
-__declspec(dllexport) const wchar_t* GetTranslationWChar(const wchar_t* text)
+DLLEXPORT const wchar_t* GetTranslationWChar(const wchar_t* text)
 {
 	return wxGetTranslation(text).wc_str();
 }
 
-__declspec(dllexport) int wxGetEventByName(const char* eventName)
+DLLEXPORT int wxGetEventByName(const char* eventName)
 {
 #define PROCESS_OWN_WXEVT(EventVarName,EventHookId) if (!strcmp(eventName,#EventVarName)){ return static_cast<int>(EventVarName); }
 #include "wxEvtHook.inl"
@@ -95,7 +95,7 @@ void FixupWxEvtIdsToMatchCemuHook()
 
 // these I added on my own since they might be useful
 
-__declspec(dllexport) void coreinitAPI_OSYieldThread()
+DLLEXPORT void coreinitAPI_OSYieldThread()
 {
 	PPCCore_switchToScheduler();
 }
diff --git a/src/config/XMLConfig.h b/src/config/XMLConfig.h
index d626484d..fd2f04c7 100644
--- a/src/config/XMLConfig.h
+++ b/src/config/XMLConfig.h
@@ -148,49 +148,41 @@ public:
 		return default_value;
 	}
 
-	template <>
 	bool value(bool default_value)
 	{
 		return m_current_element ? m_current_element->BoolText(default_value) : default_value;
 	}
 
-	template <>
 	float value(float default_value)
 	{
 		return m_current_element ? m_current_element->FloatText(default_value) : default_value;
 	}
 
-	template <>
 	double value(double default_value)
 	{
 		return m_current_element ? m_current_element->DoubleText(default_value) : default_value;
 	}
 
-	template <>
 	uint32 value(uint32 default_value)
 	{
 		return m_current_element ? m_current_element->UnsignedText(default_value) : default_value;
 	}
 
-	template <>
 	sint32 value(sint32 default_value)
 	{
 		return m_current_element ? m_current_element->IntText(default_value) : default_value;
 	}
 
-	template <>
 	uint64 value(uint64 default_value)
 	{
 		return m_current_element ? (uint64)m_current_element->Int64Text(default_value) : default_value;
 	}
 
-	template <>
 	sint64 value(sint64 default_value)
 	{
 		return m_current_element ? m_current_element->Int64Text(default_value) : default_value;
 	}
 
-	template <>
 	const char* value(const char* default_value)
 	{
 		if (m_current_element)
@@ -243,7 +235,6 @@ public:
 		set(name, value.load());
 	}
 
-	template <>
 	void set(const char* name, uint64 value)
 	{
 		set(name, (sint64)value);
@@ -300,7 +291,6 @@ public:
 		return set_attribute(name, value.GetValue());
 	}
 	
-	template <>
 	XMLConfigParser& set_attribute(const char* name, const std::string& value)
 	{
 		return set_attribute(name, value.c_str());
diff --git a/src/gui/CemuApp.cpp b/src/gui/CemuApp.cpp
index a13ba419..1ce44d4e 100644
--- a/src/gui/CemuApp.cpp
+++ b/src/gui/CemuApp.cpp
@@ -70,7 +70,7 @@ void unused_translation_dummy()
 
 
 #pragma optimize( "", off )
-DLLEXPORT _declspec(noinline) wxTopLevelWindow* wxMainWindowCreated(wxTopLevelWindow* wndPtr, uint32 magicConstant, CemuApp* appPointer)
+DLLEXPORT NOINLINE wxTopLevelWindow* wxMainWindowCreated(wxTopLevelWindow* wndPtr, uint32 magicConstant, CemuApp* appPointer)
 {
 	return wndPtr;
 }
diff --git a/src/gui/MemorySearcherTool.h b/src/gui/MemorySearcherTool.h
index 2f891062..55892f45 100644
--- a/src/gui/MemorySearcherTool.h
+++ b/src/gui/MemorySearcherTool.h
@@ -72,9 +72,6 @@ private:
 		return iss && iss.eof();
 	}
 	
-	template <>
-	bool ConvertStringToType(const char* inValue, sint8& outValue) const;
-
 	using ListType_t = std::vector<MEMPTR<void>>;
 	ListType_t SearchValues(SearchDataType type, void* ptr, uint32 size)
 	{
@@ -197,6 +194,10 @@ wxDECLARE_EVENT_TABLE();
 	bool m_clear_state = false;
 };
 
+template <>
+bool MemorySearcherTool::ConvertStringToType(const char* inValue, sint8& outValue) const;
+
+
 
 
 
diff --git a/src/gui/components/wxTitleManagerList.cpp b/src/gui/components/wxTitleManagerList.cpp
index d826c118..a68d61f1 100644
--- a/src/gui/components/wxTitleManagerList.cpp
+++ b/src/gui/components/wxTitleManagerList.cpp
@@ -941,7 +941,7 @@ wxString wxTitleManagerList::GetTitleEntryText(const TitleEntry& entry, ItemColu
 		//return wxStringFormat2("{}", entry.format);
 	}
 	default:
-		__assume(false);
+		ASSUME(false);
 	}
 	
 	return wxEmptyString;
diff --git a/src/main.cpp b/src/main.cpp
index 98f829de..1aaa6639 100644
--- a/src/main.cpp
+++ b/src/main.cpp
@@ -39,11 +39,14 @@
 #define _putenv(__s) putenv((char*)(__s))
 #endif
 
+#if BOOST_OS_WINDOWS > 0
 extern "C"
 {
-	__declspec(dllexport) int AmdPowerXpressRequestHighPerformance = 1;
-	__declspec(dllexport) DWORD NvOptimusEnablement = 0x00000001;
+	DLLEXPORT int AmdPowerXpressRequestHighPerformance = 1;
+	DLLEXPORT DWORD NvOptimusEnablement = 0x00000001;
 }
+#endif
+#endif
 
 bool _cpuExtension_SSSE3 = false;
 bool _cpuExtension_SSE4_1 = false;
@@ -51,7 +54,9 @@ bool _cpuExtension_AVX2 = false;
 
 std::atomic_bool g_isGPUInitFinished = false;
 
+#if BOOST_OS_WINDOWS > 0
 std::wstring executablePath;
+#endif
 
 bool g_cemuhook_loaded = false;
 bool IsCemuhookLoaded()
@@ -236,7 +241,7 @@ void mainEmulatorCommonInit()
 	_cpuExtension_SSSE3 = ((cpuInfo[2] >> 9) & 1) != 0;
 	_cpuExtension_SSE4_1 = ((cpuInfo[2] >> 19) & 1) != 0;
 
-	__cpuidex(cpuInfo, 0x7, 0);
+	__cpuidex1(cpuInfo, 0x7, 0);
 	_cpuExtension_AVX2 = ((cpuInfo[1] >> 5) & 1) != 0;
 
 #if BOOST_OS_WINDOWS > 0
@@ -426,7 +431,7 @@ int main(int argc, char *argv[])
 
 #pragma optimize("",off)
 
-__declspec(dllexport) void gameMeta_loadForCurrent()
+DLLEXPORT void gameMeta_loadForCurrent()
 {
 	int placeholderA = 0x11223344;
 	int placeholderB = 0x55667788;
@@ -434,7 +439,7 @@ __declspec(dllexport) void gameMeta_loadForCurrent()
 
 #pragma optimize("",on)
 
-__declspec(dllexport) uint64 gameMeta_getTitleId()
+DLLEXPORT uint64 gameMeta_getTitleId()
 {
 	return CafeSystem::GetForegroundTitleId();
 }
diff --git a/src/util/crypto/aes128.cpp b/src/util/crypto/aes128.cpp
index 17f049f3..bfd42232 100644
--- a/src/util/crypto/aes128.cpp
+++ b/src/util/crypto/aes128.cpp
@@ -759,7 +759,7 @@ void AESNI128_CBC_decryptWithExpandedKey(const unsigned char *in,
 
 void __aesni__AES128_CBC_decrypt(uint8* output, uint8* input, uint32 length, const uint8* key, const uint8* iv)
 {
-	__declspec(align(16)) uint8 expandedKey[11 * 16];
+	ALIGN(16) uint8 expandedKey[11 * 16];
 	AESNI128_KeyExpansionDecrypt(key, expandedKey);
 	if (iv)
 	{
@@ -774,7 +774,7 @@ void __aesni__AES128_CBC_decrypt(uint8* output, uint8* input, uint32 length, con
 
 void __aesni__AES128_ECB_encrypt(uint8* input, const uint8* key, uint8* output)
 {
-	__declspec(align(16)) uint8 expandedKey[11 * 16];
+	ALIGN(16) uint8 expandedKey[11 * 16];
 	AESNI128_KeyExpansionEncrypt(key, expandedKey);
 	// encrypt single ECB block
 	__m128i feedback;
diff --git a/src/util/helpers/Serializer.cpp b/src/util/helpers/Serializer.cpp
new file mode 100644
index 00000000..e9e60b0a
--- /dev/null
+++ b/src/util/helpers/Serializer.cpp
@@ -0,0 +1,162 @@
+#include "Serializer.h"
+
+template<>
+uint8 MemStreamReader::readBE()
+{
+	if (!reserveReadLength(sizeof(uint8)))
+		return 0;
+	uint8 v = m_data[m_cursorPos];
+	m_cursorPos += sizeof(uint8);
+	return v;
+}
+
+template<>
+uint16 MemStreamReader::readBE()
+{
+	if (!reserveReadLength(sizeof(uint16)))
+		return 0;
+	const uint8* p = m_data + m_cursorPos;
+	uint16 v;
+	std::memcpy(&v, p, sizeof(v));
+	v = _BE(v);
+	m_cursorPos += sizeof(uint16);
+	return v;
+}
+
+template<>
+uint32 MemStreamReader::readBE()
+{
+	if (!reserveReadLength(sizeof(uint32)))
+		return 0;
+	const uint8* p = m_data + m_cursorPos;
+	uint32 v;
+	std::memcpy(&v, p, sizeof(v));
+	v = _BE(v);
+	m_cursorPos += sizeof(uint32);
+	return v;
+}
+
+template<>
+uint64 MemStreamReader::readBE()
+{
+	if (!reserveReadLength(sizeof(uint64)))
+		return 0;
+	const uint8* p = m_data + m_cursorPos;
+	uint64 v;
+	std::memcpy(&v, p, sizeof(v));
+	v = _BE(v);
+	m_cursorPos += sizeof(uint64);
+	return v;
+}
+
+template<>
+std::string MemStreamReader::readBE()
+{
+	std::string s;
+	uint32 stringSize = readBE<uint32>();
+	if (hasError())
+		return s;
+	if (stringSize >= (32 * 1024 * 1024))
+	{
+		// out of bounds read or suspiciously large string
+		m_hasError = true;
+		return std::string();
+	}
+	s.resize(stringSize);
+	readData(s.data(), stringSize);
+	return s;
+}
+
+template<>
+uint8 MemStreamReader::readLE()
+{
+	return readBE<uint8>();
+}
+
+template<>
+uint32 MemStreamReader::readLE()
+{
+	if (!reserveReadLength(sizeof(uint32)))
+		return 0;
+	const uint8* p = m_data + m_cursorPos;
+	uint32 v;
+	std::memcpy(&v, p, sizeof(v));
+	v = _LE(v);
+	m_cursorPos += sizeof(uint32);
+	return v;
+}
+
+template<>
+uint64 MemStreamReader::readLE()
+{
+	if (!reserveReadLength(sizeof(uint64)))
+		return 0;
+	const uint8* p = m_data + m_cursorPos;
+	uint64 v;
+	std::memcpy(&v, p, sizeof(v));
+	v = _LE(v);
+	m_cursorPos += sizeof(uint64);
+	return v;
+}
+
+template<>
+void MemStreamWriter::writeBE<uint64>(const uint64& v)
+{
+	m_buffer.resize(m_buffer.size() + 8);
+	uint8* p = m_buffer.data() + m_buffer.size() - 8;
+	uint64 tmp = _BE(v);
+	std::memcpy(p, &tmp, sizeof(tmp));
+}
+
+template<>
+void MemStreamWriter::writeBE<uint32>(const uint32& v)
+{
+	m_buffer.resize(m_buffer.size() + 4);
+	uint8* p = m_buffer.data() + m_buffer.size() - 4;
+	uint32 tmp = _BE(v);
+	std::memcpy(p, &tmp, sizeof(tmp));
+}
+
+
+template<>
+void MemStreamWriter::writeBE<uint16>(const uint16& v)
+{
+	m_buffer.resize(m_buffer.size() + 2);
+	uint8* p = m_buffer.data() + m_buffer.size() - 2;
+	uint16 tmp = _BE(v);
+	std::memcpy(p, &tmp, sizeof(tmp));
+}
+
+
+template<>
+void MemStreamWriter::writeBE<uint8>(const uint8& v)
+{
+	m_buffer.emplace_back(v);
+}
+
+
+template<>
+void MemStreamWriter::writeBE<std::string>(const std::string& v)
+{
+	writeBE<uint32>((uint32)v.size());
+	writeData(v.data(), v.size());
+}
+
+template<>
+void MemStreamWriter::writeLE<uint64>(const uint64& v)
+{
+	m_buffer.resize(m_buffer.size() + 8);
+	uint8* p = m_buffer.data() + m_buffer.size() - 8;
+	uint64 tmp = _LE(v);
+	std::memcpy(p, &tmp, sizeof(tmp));
+}
+
+
+template<>
+void MemStreamWriter::writeLE<uint32>(const uint32& v)
+{
+	m_buffer.resize(m_buffer.size() + 4);
+	uint8* p = m_buffer.data() + m_buffer.size() - 4;
+	uint32 tmp = _LE(v);
+	std::memcpy(p, &tmp, sizeof(tmp));
+}
\ No newline at end of file
diff --git a/src/util/helpers/Serializer.h b/src/util/helpers/Serializer.h
index a15adfad..3d82cbeb 100644
--- a/src/util/helpers/Serializer.h
+++ b/src/util/helpers/Serializer.h
@@ -11,97 +11,6 @@ public:
 	template<typename T> T readBE();
 	template<typename T> T readLE();
 
-	template<> uint8 readBE()
-	{
-		if (!reserveReadLength(sizeof(uint8)))
-			return 0;
-		uint8 v = m_data[m_cursorPos];
-		m_cursorPos += sizeof(uint8);
-		return v;
-	}
-
-	template<> uint16 readBE()
-	{
-		if (!reserveReadLength(sizeof(uint16)))
-			return 0;
-		const uint8* p = m_data + m_cursorPos;
-		uint16 v;
-		std::memcpy(&v, p, sizeof(v));
-		v = _BE(v);
-		m_cursorPos += sizeof(uint16);
-		return v;
-	}
-
-	template<> uint32 readBE()
-	{
-		if (!reserveReadLength(sizeof(uint32)))
-			return 0;
-		const uint8* p = m_data + m_cursorPos;
-		uint32 v;
-		std::memcpy(&v, p, sizeof(v));
-		v = _BE(v);
-		m_cursorPos += sizeof(uint32);
-		return v;
-	}
-
-	template<> uint64 readBE()
-	{
-		if (!reserveReadLength(sizeof(uint64)))
-			return 0;
-		const uint8* p = m_data + m_cursorPos;
-		uint64 v;
-		std::memcpy(&v, p, sizeof(v));
-		v = _BE(v);
-		m_cursorPos += sizeof(uint64);
-		return v;
-	}
-
-	template<> std::string readBE()
-	{
-		std::string s;
-		uint32 stringSize = readBE<uint32>();
-		if (hasError())
-			return s;
-		if (stringSize >= (32 * 1024 * 1024))
-		{
-			// out of bounds read or suspiciously large string
-			m_hasError = true;
-			return std::string();
-		}
-		s.resize(stringSize);
-		readData(s.data(), stringSize);
-		return s;
-	}
-
-	template<> uint8 readLE()
-	{
-		return readBE<uint8>();
-	}
-
-	template<> uint32 readLE()
-	{
-		if (!reserveReadLength(sizeof(uint32)))
-			return 0;
-		const uint8* p = m_data + m_cursorPos;
-		uint32 v;
-		std::memcpy(&v, p, sizeof(v));
-		v = _LE(v);
-		m_cursorPos += sizeof(uint32);
-		return v;
-	}
-
-	template<> uint64 readLE()
-	{
-		if (!reserveReadLength(sizeof(uint64)))
-			return 0;
-		const uint8* p = m_data + m_cursorPos;
-		uint64 v;
-		std::memcpy(&v, p, sizeof(v));
-		v = _LE(v);
-		m_cursorPos += sizeof(uint64);
-		return v;
-	}
-
 	template<typename T>
 	std::vector<T> readPODVector()
 	{
@@ -224,6 +133,7 @@ private:
 	bool m_hasError{ false };
 };
 
+
 class MemStreamWriter
 {
 public:
@@ -243,67 +153,8 @@ public:
 	}
 
 	template<typename T> void writeBE(const T& v);
-
-	template<>
-	void writeBE<uint64>(const uint64& v)
-	{
-		m_buffer.resize(m_buffer.size() + 8);
-		uint8* p = m_buffer.data() + m_buffer.size() - 8;
-		uint64 tmp = _BE(v);
-		std::memcpy(p, &tmp, sizeof(tmp));
-	}
-
-	template<>
-	void writeBE<uint32>(const uint32& v)
-	{
-		m_buffer.resize(m_buffer.size() + 4);
-		uint8* p = m_buffer.data() + m_buffer.size() - 4;
-		uint32 tmp = _BE(v);
-		std::memcpy(p, &tmp, sizeof(tmp));
-	}
-
-	template<>
-	void writeBE<uint16>(const uint16& v)
-	{
-		m_buffer.resize(m_buffer.size() + 2);
-		uint8* p = m_buffer.data() + m_buffer.size() - 2;
-		uint16 tmp = _BE(v);
-		std::memcpy(p, &tmp, sizeof(tmp));
-	}
-
-	template<>
-	void writeBE<uint8>(const uint8& v)
-	{
-		m_buffer.emplace_back(v);
-	}
-
-	template<>
-	void writeBE<std::string>(const std::string& v)
-	{
-		writeBE<uint32>((uint32)v.size());
-		writeData(v.data(), v.size());
-	}
-
 	template<typename T> void writeLE(const T& v);
 
-	template<>
-	void writeLE<uint64>(const uint64& v)
-	{
-		m_buffer.resize(m_buffer.size() + 8);
-		uint8* p = m_buffer.data() + m_buffer.size() - 8;
-		uint64 tmp = _LE(v);
-		std::memcpy(p, &tmp, sizeof(tmp));
-	}
-
-	template<>
-	void writeLE<uint32>(const uint32& v)
-	{
-		m_buffer.resize(m_buffer.size() + 4);
-		uint8* p = m_buffer.data() + m_buffer.size() - 4;
-		uint32 tmp = _LE(v);
-		std::memcpy(p, &tmp, sizeof(tmp));
-	}
-
 	template<typename T>
 	void writePODVector(const std::vector<T>& v)
 	{
@@ -328,6 +179,7 @@ private:
 	std::vector<uint8> m_buffer;
 };
 
+
 class SerializerHelper 
 {
 public:
diff --git a/src/util/helpers/ringbuffer.h b/src/util/helpers/ringbuffer.h
index ae036efd..a4b6ab79 100644
--- a/src/util/helpers/ringbuffer.h
+++ b/src/util/helpers/ringbuffer.h
@@ -6,7 +6,8 @@ template<typename T, uint32 elements, typename P = uint32>
 class RingBuffer
 {
 public:
-	RingBuffer<T, elements, P>();
+	//RingBuffer<T, elements, P>();
+	RingBuffer();
 
 	bool Push(const T& v);