Merge pull request #1800 from randomstuff/jit-register

Add support of more JIT-compiled code for profiling

Source/Core/Common/JitRegister.cpp

@@ -71,19 +71,15 @@ void Shutdown()
 		s_perf_map_file.Close();
 }
 
-void Register(const void* base_address, u32 code_size,
-	const char* name, u32 original_address)
+void RegisterV(const void* base_address, u32 code_size,
+	const char* format, va_list args)
 {
 #if !(defined USE_OPROFILE && USE_OPROFILE) && !defined(USE_VTUNE)
 	if (!s_perf_map_file.IsOpen())
 		return;
 #endif
 
-	std::string symbol_name;
-	if (original_address)
-		symbol_name = StringFromFormat("%s_%x", name, original_address);
-	else
-		symbol_name = name;
+	std::string symbol_name = StringFromFormatV(format, args);
 
 #if defined USE_OPROFILE && USE_OPROFILE
 	op_write_native_code(s_agent, symbol_name.data(), (u64)base_address,

Source/Core/Common/JitRegister.h

@@ -3,6 +3,7 @@
 // Refer to the license.txt file included.
 
 #pragma once
+#include <stdarg.h>
 #include "Common/CommonTypes.h"
 
 namespace JitRegister
@@ -10,7 +11,26 @@ namespace JitRegister
 
 void Init();
 void Shutdown();
-void Register(const void* base_address, u32 code_size,
-	const char* name, u32 original_address=0);
+void RegisterV(const void* base_address, u32 code_size,
+	const char* format, va_list args);
+
+inline void Register(const void* base_address, u32 code_size,
+	const char* format, ...)
+{
+	va_list args;
+	va_start(args, format);
+	RegisterV(base_address, code_size, format, args);
+	va_end(args);
+}
+
+inline void Register(const void* start, const void* end,
+	const char* format, ...)
+{
+	va_list args;
+	va_start(args, format);
+	u32 code_size = (u32) ((const char*) end - (const char*) start);
+	RegisterV(start, code_size, format, args);
+	va_end(args);
+}
 
 }

Source/Core/Common/StringUtil.cpp

@@ -95,28 +95,30 @@ bool CharArrayFromFormatV(char* out, int outsize, const char* format, va_list ar
 std::string StringFromFormat(const char* format, ...)
 {
 	va_list args;
+	va_start(args, format);
+	std::string res = StringFromFormatV(format, args);
+	va_end(args);
+	return std::move(res);
+}
+
+std::string StringFromFormatV(const char* format, va_list args)
+{
 	char *buf = nullptr;
 #ifdef _WIN32
-	int required = 0;
-
-	va_start(args, format);
-	required = _vscprintf(format, args);
+	int required = _vscprintf(format, args);
 	buf = new char[required + 1];
 	CharArrayFromFormatV(buf, required + 1, format, args);
-	va_end(args);
 
 	std::string temp = buf;
 	delete[] buf;
 #else
-	va_start(args, format);
 	if (vasprintf(&buf, format, args) < 0)
 		ERROR_LOG(COMMON, "Unable to allocate memory for string");
-	va_end(args);
 
 	std::string temp = buf;
 	free(buf);
 #endif
-	return temp;
+	return std::move(temp);
 }
 
 // For Debugging. Read out an u8 array.

Source/Core/Common/StringUtil.h

@@ -13,6 +13,8 @@
 
 #include "Common/Common.h"
 
+std::string StringFromFormatV(const char* format, va_list args);
+
 std::string StringFromFormat(const char* format, ...)
 #if !defined _WIN32
 // On compilers that support function attributes, this gives StringFromFormat

Source/Core/Core/PowerPC/Jit64/JitAsm.cpp

@@ -2,6 +2,7 @@
 // Licensed under GPLv2
 // Refer to the license.txt file included.
 
+#include "Common/JitRegister.h"
 #include "Common/MemoryUtil.h"
 
 #include "Core/PowerPC/Jit64/Jit.h"
@@ -200,6 +201,8 @@ void Jit64AsmRoutineManager::Generate()
 	ABI_PopRegistersAndAdjustStack(ABI_ALL_CALLEE_SAVED, 8, 16);
 	RET();
 
+	JitRegister::Register(enterCode, GetCodePtr(), "JIT_Loop");
+
 	GenerateCommon();
 }
 

Source/Core/Core/PowerPC/Jit64Common/Jit64AsmCommon.cpp

@@ -2,6 +2,7 @@
 // Licensed under GPLv2
 // Refer to the license.txt file included.
 
+#include "Common/JitRegister.h"
 #include "Common/MathUtil.h"
 #include "Common/x64ABI.h"
 #include "Common/x64Emitter.h"
@@ -20,6 +21,8 @@ using namespace Gen;
 
 void CommonAsmRoutines::GenFifoWrite(int size)
 {
+	const void* start = GetCodePtr();
+
 	// Assume value in RSCRATCH
 	u32 gather_pipe = (u32)(u64)GPFifo::m_gatherPipe;
 	_assert_msg_(DYNA_REC, gather_pipe <= 0x7FFFFFFF, "Gather pipe not in low 2GB of memory!");
@@ -28,10 +31,14 @@ void CommonAsmRoutines::GenFifoWrite(int size)
 	ADD(32, R(RSCRATCH2), Imm8(size >> 3));
 	MOV(32, M(&GPFifo::m_gatherPipeCount), R(RSCRATCH2));
 	RET();
+
+	JitRegister::Register(start, GetCodePtr(), "JIT_FifoWrite_%i", size);
 }
 
 void CommonAsmRoutines::GenFrsqrte()
 {
+	const void* start = GetCodePtr();
+
 	// Assume input in XMM0.
 	// This function clobbers all three RSCRATCH.
 	MOVQ_xmm(R(RSCRATCH), XMM0);
@@ -91,10 +98,14 @@ void CommonAsmRoutines::GenFrsqrte()
 	ABI_CallFunction((void *)&MathUtil::ApproximateReciprocalSquareRoot);
 	ABI_PopRegistersAndAdjustStack(QUANTIZED_REGS_TO_SAVE, 8);
 	RET();
+
+	JitRegister::Register(start, GetCodePtr(), "JIT_Frsqrte");
 }
 
 void CommonAsmRoutines::GenFres()
 {
+	const void* start = GetCodePtr();
+
 	// Assume input in XMM0.
 	// This function clobbers all three RSCRATCH.
 	MOVQ_xmm(R(RSCRATCH), XMM0);
@@ -149,10 +160,14 @@ void CommonAsmRoutines::GenFres()
 	ABI_CallFunction((void *)&MathUtil::ApproximateReciprocal);
 	ABI_PopRegistersAndAdjustStack(QUANTIZED_REGS_TO_SAVE, 8);
 	RET();
+
+	JitRegister::Register(start, GetCodePtr(), "JIT_Fres");
 }
 
 void CommonAsmRoutines::GenMfcr()
 {
+	const void* start = GetCodePtr();
+
 	// Input: none
 	// Output: RSCRATCH
 	// This function clobbers all three RSCRATCH.
@@ -187,6 +202,8 @@ void CommonAsmRoutines::GenMfcr()
 		OR(32, R(dst), MScaled(cr_val, SCALE_4, (u32)(u64)m_flagTable));
 	}
 	RET();
+
+	JitRegister::Register(start, GetCodePtr(), "JIT_Mfcr");
 }
 
 // Safe + Fast Quantizers, originally from JITIL by magumagu
@@ -207,6 +224,8 @@ static const float GC_ALIGNED16(m_m128) = -128.0f;
 // See comment in header for in/outs.
 void CommonAsmRoutines::GenQuantizedStores()
 {
+	const void* start = GetCodePtr();
+
 	const u8* storePairedIllegal = AlignCode4();
 	UD2();
 
@@ -305,6 +324,8 @@ void CommonAsmRoutines::GenQuantizedStores()
 
 	RET();
 
+	JitRegister::Register(start, GetCodePtr(), "JIT_QuantizedStore");
+
 	pairedStoreQuantized = reinterpret_cast<const u8**>(const_cast<u8*>(AlignCode16()));
 	ReserveCodeSpace(8 * sizeof(u8*));
 
@@ -321,6 +342,8 @@ void CommonAsmRoutines::GenQuantizedStores()
 // See comment in header for in/outs.
 void CommonAsmRoutines::GenQuantizedSingleStores()
 {
+	const void* start = GetCodePtr();
+
 	const u8* storeSingleIllegal = AlignCode4();
 	UD2();
 
@@ -368,6 +391,8 @@ void CommonAsmRoutines::GenQuantizedSingleStores()
 	SafeWriteRegToReg(RSCRATCH, RSCRATCH_EXTRA, 16, 0, QUANTIZED_REGS_TO_SAVE, SAFE_LOADSTORE_NO_PROLOG | SAFE_LOADSTORE_NO_FASTMEM);
 	RET();
 
+	JitRegister::Register(start, GetCodePtr(), "JIT_QuantizedSingleStore");
+
 	singleStoreQuantized = reinterpret_cast<const u8**>(const_cast<u8*>(AlignCode16()));
 	ReserveCodeSpace(8 * sizeof(u8*));
 
@@ -383,6 +408,8 @@ void CommonAsmRoutines::GenQuantizedSingleStores()
 
 void CommonAsmRoutines::GenQuantizedLoads()
 {
+	const void* start = GetCodePtr();
+
 	const u8* loadPairedIllegal = AlignCode4();
 	UD2();
 
@@ -578,6 +605,9 @@ void CommonAsmRoutines::GenQuantizedLoads()
 	UNPCKLPS(XMM0, M(m_one));
 	RET();
 
+
+	JitRegister::Register(start, GetCodePtr(), "JIT_QuantizedLoad");
+
 	pairedLoadQuantized = reinterpret_cast<const u8**>(const_cast<u8*>(AlignCode16()));
 	ReserveCodeSpace(16 * sizeof(u8*));
 

Source/Core/Core/PowerPC/JitCommon/JitCache.cpp

@@ -144,7 +144,7 @@ using namespace Gen;
 		}
 
 		JitRegister::Register(blockCodePointers[block_num], b.codeSize,
-			"JIT_PPC", b.originalAddress);
+			"JIT_PPC_%08x", b.originalAddress);
 	}
 
 	const u8 **JitBaseBlockCache::GetCodePointers()

Source/Core/Core/PowerPC/JitCommon/TrampolineCache.cpp

@@ -6,6 +6,7 @@
 #include <string>
 
 #include "Common/CommonTypes.h"
+#include "Common/JitRegister.h"
 #include "Common/StringUtil.h"
 #include "Common/x64ABI.h"
 #include "Core/HW/Memmap.h"
@@ -95,6 +96,8 @@ const u8* TrampolineCache::GenerateReadTrampoline(const InstructionInfo &info, B
 		MOVZX(dataRegSize, info.operandSize * 8, dataReg, R(ABI_RETURN));
 
 	JMP(returnPtr, true);
+
+	JitRegister::Register(trampoline, GetCodePtr(), "JIT_ReadTrampoline");
 	return trampoline;
 }
 
@@ -172,5 +175,6 @@ const u8* TrampolineCache::GenerateWriteTrampoline(const InstructionInfo &info,
 	}
 	JMP(returnPtr, true);
 
+	JitRegister::Register(trampoline, GetCodePtr(), "JIT_WriteTrampoline_%x", pc);
 	return trampoline;
 }

Source/Core/VideoCommon/VertexLoaderX64.cpp

@@ -33,7 +33,7 @@ VertexLoaderX64::VertexLoaderX64(const TVtxDesc& vtx_desc, const VAT& vtx_att):
 
 	std::string name;
 	AppendToString(&name);
-	JitRegister::Register(region, (u32)(GetCodePtr() - region), name.c_str());
+	JitRegister::Register(region, GetCodePtr(), name.c_str());
 }
 
 OpArg VertexLoaderX64::GetVertexAddr(int array, u64 attribute)