vtune: plug PCSX2 core + add missing profiling (VU/VIF/TLB)

Doesn't fully work yet
* Unknown stack frame
* Outside any known module

Potential root cause:
* Nvidia driver
* VU code as ebp is required for emulation so likely no frame

build.sh

@@ -114,6 +114,7 @@ for ARG in "$@"; do
             echo "--clang-tidy    : Do a clang-tidy analysis. Results can be found in build directory"
             echo "--cppcheck      : Do a cppcheck analysis. Results can be found in build directory"
             echo "--coverity      : Do a build for coverity"
+            echo "--vtune         : Plug GSdx with VTUNE"
 
             exit 1
     esac

common/include/Utilities/Perf.h

@@ -53,4 +53,5 @@ extern InfoVector any;
 extern InfoVector ee;
 extern InfoVector iop;
 extern InfoVector vu;
+extern InfoVector vif;
 }

common/src/Utilities/CMakeLists.txt

@@ -107,5 +107,10 @@ set(UtilitiesFinalLibs
 	${wxWidgets_LIBRARIES}
 )
 
+if(USE_VTUNE)
+	set(UtilitiesFinalFlags ${UtilitiesFinalFlags} -DENABLE_VTUNE)
+	include_directories("$ENV{VTUNE_AMPLIFIER_XE_2016_DIR}/include")
+endif()
+
 add_pcsx2_lib(${Output} "${UtilitiesFinalSources}" "${UtilitiesFinalLibs}" "${UtilitiesFinalFlags}")
 add_pcsx2_lib(${Output}_NO_TLS "${UtilitiesFinalSources}" "${UtilitiesFinalLibs}" "${UtilitiesFinalFlags} -DPCSX2_THREAD_LOCAL=0")

common/src/Utilities/Perf.cpp

@@ -17,9 +17,20 @@
 
 #include "Perf.h"
 
+#ifdef __unix__
+#include "unistd.h"
+#endif
+
 //#define ProfileWithPerf
 #define MERGE_BLOCK_RESULT
 
+#ifdef ENABLE_VTUNE
+	#include "jitprofiling.h"
+
+	#ifdef _WIN32
+	#pragma comment(lib, "jitprofiling.lib")
+	#endif
+#endif
 
 namespace Perf
 {
@@ -28,9 +39,10 @@ InfoVector any("");
 InfoVector ee("EE");
 InfoVector iop("IOP");
 InfoVector vu("VU");
+InfoVector vif("VIF");
 
 // Perf is only supported on linux
-#if defined(__linux__) && defined(ProfileWithPerf)
+#if defined(__linux__) && (defined(ProfileWithPerf) || defined(ENABLE_VTUNE))
 
 ////////////////////////////////////////////////////////////////////////////////
 // Implementation of the Info object
@@ -85,6 +97,25 @@ void InfoVector::map(uptr x86, u32 size, const char *symbol)
     if (size < 8 * _1kb)
         m_v.emplace_back(x86, size, symbol);
 #endif
+
+#ifdef ENABLE_VTUNE
+	// mapping the full recompiler will blow up VTUNE
+	if (size < _16kb) {
+		fprintf(stderr, "map %s: %p size %d\n", symbol, (void*)x86, size);
+		std::string name = std::string(symbol);
+
+		iJIT_Method_Load ml;
+
+		memset(&ml, 0, sizeof(ml));
+
+		ml.method_id = iJIT_GetNewMethodID();
+		ml.method_name = (char*)name.c_str();
+		ml.method_load_address = (void*)x86;
+		ml.method_size = size;
+
+		iJIT_NotifyEvent(iJVM_EVENT_TYPE_METHOD_LOAD_FINISHED, &ml);
+	}
+#endif
 }
 
 void InfoVector::map(uptr x86, u32 size, u32 pc)
@@ -92,6 +123,22 @@ void InfoVector::map(uptr x86, u32 size, u32 pc)
 #ifndef MERGE_BLOCK_RESULT
     m_v.emplace_back(x86, size, m_prefix, pc);
 #endif
+
+#ifdef ENABLE_VTUNE
+	std::string name = std::string(m_prefix) + "_" + std::to_string(pc);
+	//fprintf(stderr, "map %s: %p size %d\n", name.c_str(), (void*)x86, size);
+
+	iJIT_Method_Load ml;
+
+	memset(&ml, 0, sizeof(ml));
+
+	ml.method_id = iJIT_GetNewMethodID();
+	ml.method_name = (char*)name.c_str();
+	ml.method_load_address = (void*)x86;
+	ml.method_size = size;
+
+	iJIT_NotifyEvent(iJVM_EVENT_TYPE_METHOD_LOAD_FINISHED, &ml);
+#endif
 }
 
 void InfoVector::reset()

pcsx2/CMakeLists.txt

@@ -681,6 +681,16 @@ foreach(res_file IN ITEMS
     add_custom_command(OUTPUT  "${res_bin}/${res_file}.h" COMMAND perl ${CMAKE_SOURCE_DIR}/linux_various/hex2h.pl "${res_src}/${res_file}.png" "${res_bin}/${res_file}" )
 endforeach()
 
+if(USE_VTUNE)
+	set(pcsx2FinalFlags ${pcsx2FinalFlags} -DENABLE_VTUNE)
+	include_directories("$ENV{VTUNE_AMPLIFIER_XE_2016_DIR}/include")
+	if(${PCSX2_TARGET_ARCHITECTURES} MATCHES "i386")
+		set(pcsx2FinalLibs ${pcsx2FinalLibs} $ENV{VTUNE_AMPLIFIER_XE_2016_DIR}/lib32/libjitprofiling.a)
+	else()
+		set(pcsx2FinalLibs ${pcsx2FinalLibs} $ENV{VTUNE_AMPLIFIER_XE_2016_DIR}/lib64/libjitprofiling.a)
+	endif()
+endif()
+
 # Suppress all the system-specific predefined macros outside the reserved namespace.
 # Needed when stringifying macros.
 # See issue: 1233

pcsx2/x86/iR3000A.cpp

@@ -170,7 +170,11 @@ static DynGenFunc* _DynGen_EnterRecompiledCode()
 	u8* retval = xGetPtr();
 
 	{ // Properly scope the frame prologue/epilogue
+#ifdef ENABLE_VTUNE
+		xScopedStackFrame frame(true);
+#else
 		xScopedStackFrame frame(IsDevBuild);
+#endif
 
 		xJMP((void*)iopDispatcherReg);
 

pcsx2/x86/ix86-32/iR5900-32.cpp

@@ -393,7 +393,11 @@ static DynGenFunc* _DynGen_EnterRecompiledCode()
 	u8* retval = xGetAlignedCallTarget();
 
 	{ // Properly scope the frame prologue/epilogue
+#ifdef ENABLE_VTUNE
+		xScopedStackFrame frame(true);
+#else
 		xScopedStackFrame frame(IsDevBuild);
+#endif
 
 		xJMP((void*)DispatcherReg);
 

pcsx2/x86/ix86-32/recVTLB.cpp

@@ -21,6 +21,7 @@
 
 #include "iCore.h"
 #include "iR5900.h"
+#include "Utilities/Perf.h"
 
 using namespace vtlb_private;
 using namespace x86Emitter;
@@ -333,6 +334,8 @@ void vtlb_dynarec_init()
 	}
 
 	HostSys::MemProtectStatic( m_IndirectDispatchers, PageAccess_ExecOnly() );
+
+	Perf::any.map((uptr)m_IndirectDispatchers, __pagesize, "TLB Dispatcher");
 }
 
 //////////////////////////////////////////////////////////////////////////////////////////

pcsx2/x86/microVU.h

@@ -35,6 +35,7 @@ using namespace x86Emitter;
 #include "microVU_Misc.h"
 #include "microVU_IR.h"
 #include "microVU_Profiler.h"
+#include "Utilities/Perf.h"
 
 struct microBlockLink {
 	microBlock		block;

pcsx2/x86/microVU_Compile.inl

@@ -535,7 +535,7 @@ void mVUSaveFlags(microVU& mVU,microFlagCycles &mFC, microFlagCycles &mFCBackup)
 	mVUsetFlags(mVU, mFCBackup);	   // Sets Up Flag instances
 }
 void* mVUcompile(microVU& mVU, u32 startPC, uptr pState) {
-	
+
 	microFlagCycles mFC;
 	u8*				thisPtr  = x86Ptr;
 	const u32		endCount = (((microRegInfo*)pState)->blockType) ? 1 : (mVU.microMemSize / 8);
@@ -597,7 +597,7 @@ void* mVUcompile(microVU& mVU, u32 startPC, uptr pState) {
 	// Fix up vi15 const info for propagation through blocks
 	mVUregs.vi15  = (doConstProp && mVUconstReg[15].isValid) ? (u16)mVUconstReg[15].regValue : 0;
 	mVUregs.vi15v = (doConstProp && mVUconstReg[15].isValid) ? 1 : 0;
-		
+
 	mVUsetFlags(mVU, mFC);           // Sets Up Flag instances
 	mVUoptimizePipeState(mVU);       // Optimize the End Pipeline State for nicer Block Linking
 	mVUdebugPrintBlocks(mVU, false); // Prints Start/End PC of blocks executed, for debugging...
@@ -653,10 +653,13 @@ void* mVUcompile(microVU& mVU, u32 startPC, uptr pState) {
 		}
 	}
 	if ((x == endCount) && (x!=1)) { Console.Error("microVU%d: Possible infinite compiling loop!", mVU.index); }
-	
+
 	// E-bit End
 	mVUsetupRange(mVU, xPC-8, false);
 	mVUendProgram(mVU, &mFC, 1);
+
+	Perf::vu.map((uptr)thisPtr, x86Ptr - thisPtr, startPC);
+
 	return thisPtr;
 }
 

pcsx2/x86/newVif_Dynarec.cpp

@@ -20,6 +20,7 @@
 #include "PrecompiledHeader.h"
 #include "newVif_UnpackSSE.h"
 #include "MTVU.h"
+#include "Utilities/Perf.h"
 
 void dVifReserve(int idx) {
 	if(!nVif[idx].recReserve)
@@ -352,6 +353,8 @@ _vifT __fi void dVifUnpack(const u8* data, bool isFill) {
 	v.block.startPtr = (uptr)xGetAlignedCallTarget();
 	v.vifBlocks->add(v.block);
 	VifUnpackSSE_Dynarec(v, v.block).CompileRoutine();
+
+	Perf::vif.map((uptr)v.recWritePtr, xGetPtr() - v.recWritePtr, v.block.upkType /* FIXME ideally a key*/);
 	nVif[idx].recWritePtr = xGetPtr();
 
 	dVifRecLimit(idx);

plugins/GSdx/CMakeLists.txt

@@ -206,8 +206,11 @@ endif()
 if(USE_VTUNE)
     set(GSdxFinalFlags ${GSdxFinalFlags} -DENABLE_VTUNE)
     include_directories("$ENV{VTUNE_AMPLIFIER_XE_2016_DIR}/include")
-    set(GSdxFinalLibs ${GSdxFinalLibs} $ENV{VTUNE_AMPLIFIER_XE_2016_DIR}/lib64/libjitprofiling.a)
+	if(${PCSX2_TARGET_ARCHITECTURES} MATCHES "i386")
-    set(GSdxFinalLibs ${GSdxFinalLibs} $ENV{VTUNE_AMPLIFIER_XE_2016_DIR}/lib32/libjitprofiling.a)
+		set(GSdxFinalLibs ${GSdxFinalLibs} $ENV{VTUNE_AMPLIFIER_XE_2016_DIR}/lib32/libjitprofiling.a)
+	else()
+		set(GSdxFinalLibs ${GSdxFinalLibs} $ENV{VTUNE_AMPLIFIER_XE_2016_DIR}/lib64/libjitprofiling.a)
+	endif()
 endif()
 
 # Generate Glsl header file. Protect with REBUILD_SHADER to avoid build-dependency on PERL