cemu-project · Crementif · Aug 28, 2022 · Aug 25, 2022 · Aug 25, 2022 · Aug 26, 2022
diff --git a/dependencies/ih264d/decoder/x86/ih264d_function_selector.c b/dependencies/ih264d/decoder/x86/ih264d_function_selector.c
@@ -77,7 +77,7 @@ void ih264d_init_function_ptr(dec_struct_t *ps_codec)
     }
 }
 
-#ifdef __clang__
+#ifdef __GNUC__
 #include <cpuid.h>
 
 void __cpuid2(signed int* cpuInfo, unsigned int level)

diff --git a/src/CMakeLists.txt b/src/CMakeLists.txt
@@ -25,8 +25,7 @@ elseif(UNIX)
         add_definitions(-DVK_USE_PLATFORM_XCB_KHR)
     endif()
     add_definitions(-fms-extensions)
-    add_definitions(-fms-compatibility-version=19.14)
-    add_definitions(-fdelayed-template-parsing)
+    add_definitions(-fpermissive)
     add_definitions(-maes)
     # warnings
     add_compile_options(-Wno-switch -Wno-ignored-attributes -Wno-deprecated-enum-enum-conversion -Wno-ambiguous-reversed-operator)

diff --git a/src/Cafe/Account/Account.cpp b/src/Cafe/Account/Account.cpp
@@ -416,7 +416,7 @@ fs::path Account::GetFileName(uint32 persistent_id)
 	if (persistent_id < kMinPersistendId)
 		throw std::invalid_argument(fmt::format("persistent id {:#x} is invalid", persistent_id));
 
-	return CemuApp::GetMLCPath(fmt::format(L"usr\\save\\system\\act\\{:08x}\\account.dat", persistent_id)).ToStdWstring();
+	return CemuApp::GetMLCPath(fmt::format(L"usr/save/system/act/{:08x}/account.dat", persistent_id)).ToStdWstring();
 }
 
 OnlineValidator Account::ValidateOnlineFiles() const

diff --git 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()
diff --git a/src/Cafe/CafeSystem.cpp b/src/Cafe/CafeSystem.cpp
@@ -856,4 +856,4 @@ namespace CafeSystem
 		return currentUpdatedApplicationHash;
 	}
 
-}
+}
diff --git a/src/Cafe/GameProfile/GameProfile.cpp b/src/Cafe/GameProfile/GameProfile.cpp
@@ -368,4 +368,4 @@ void GameProfile::Reset()
 	// controller settings
 	for (auto& profile : m_controllerProfile)
 		profile.reset();
-}
+}
diff --git a/src/Cafe/GraphicPack/GraphicPack.cpp b/src/Cafe/GraphicPack/GraphicPack.cpp
@@ -124,4 +124,4 @@ void graphicPack_activateForCurrentTitle(uint64 titleId)
 			}
 		}
 	}
-}
+}
diff --git a/src/Cafe/GraphicPack/GraphicPack2.cpp b/src/Cafe/GraphicPack/GraphicPack2.cpp
@@ -1129,4 +1129,4 @@ std::vector<std::pair<MPTR, MPTR>> GraphicPack2::GetActiveRAMMappings()
 			return a.first < b.first;
 		});
 	return v;
-}
+}
diff --git a/src/Cafe/GraphicPack/GraphicPack2.h b/src/Cafe/GraphicPack/GraphicPack2.h
@@ -320,4 +320,4 @@ std::vector<T> GraphicPack2::ParseList(const ExpressionParser& parser, IniParser
 	}
 
 	return result;
-}
+}
diff --git 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/PPCTimer.cpp b/src/Cafe/HW/Espresso/PPCTimer.cpp
@@ -7,9 +7,11 @@
 
 #if BOOST_OS_LINUX || BOOST_OS_MACOS
 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.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];
@@ -396,4 +396,4 @@ extern bool hasAVXSupport;
 
 void PPCRecompiler_recompileIfUnvisited(uint32 enterAddress);
 void PPCRecompiler_attemptEnter(struct PPCInterpreter_t* hCPU, uint32 enterAddress);
-void PPCRecompiler_attemptEnterWithoutRecompile(struct PPCInterpreter_t* hCPU, uint32 enterAddress);
+void PPCRecompiler_attemptEnterWithoutRecompile(struct PPCInterpreter_t* hCPU, uint32 enterAddress);
diff --git 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));
@@ -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
@@ -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
@@ -425,4 +425,4 @@ void LatteTC_UnloadAllTextures()
 			LatteTexture_Delete(itr);
 	}
 	LatteRenderTarget_unloadAll();
-}
+}
diff --git 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
@@ -1463,7 +1463,7 @@ void OpenGLRenderer::shader_bind(RendererShader* shader)
 		prevGeometryShaderProgram = program;
 		break;
 	default:
-		__assume(false);
+		ASSUME(false);
 	}
 
 	catchOpenGLError();
@@ -1494,7 +1494,7 @@ 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
@@ -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
@@ -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
@@ -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;
-}
+}
diff --git a/src/Cafe/HW/Latte/Renderer/Vulkan/VulkanRenderer.cpp b/src/Cafe/HW/Latte/Renderer/Vulkan/VulkanRenderer.cpp
@@ -3412,7 +3412,7 @@ 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
@@ -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
@@ -533,4 +533,4 @@ namespace MMU
 	}
 
 
-}
+}
diff --git a/src/Cafe/OS/RPL/rpl.h b/src/Cafe/OS/RPL/rpl.h
@@ -52,4 +52,4 @@ uint32 RPLLoader_MakePPCCallable(void(*ppcCallableExport)(struct PPCInterpreter_
 
 // elf loader
 
-uint32 ELF_LoadFromMemory(uint8* elfData, sint32 size, const char* name);
+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
@@ -217,4 +217,4 @@ void osLib_load()
 	swkbd::load();
 	camera::load();
 	procui_load();
-}
+}
diff --git 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
@@ -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
@@ -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
@@ -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;
-Original file line number
+Diff line change
@@ Expand Up / @@ -856,4 +856,4 @@ namespace CafeSystem @@
     		return currentUpdatedApplicationHash;
     	}
-    }
+    }
Original file line number	Diff line number	Diff line change
Expand Up		@@ -52,4 +52,4 @@ uint32 RPLLoader_MakePPCCallable(void(*ppcCallableExport)(struct PPCInterpreter_

		// elf loader

		uint32 ELF_LoadFromMemory(uint8* elfData, sint32 size, const char* name);
		uint32 ELF_LoadFromMemory(uint8* elfData, sint32 size, const char* name);