SPU JIT WIP

asmjit.vcxproj

@@ -92,12 +92,12 @@
   </ImportGroup>
   <PropertyGroup Label="UserMacros" />
   <PropertyGroup Condition="'$(Configuration)|$(Platform)'=='Debug|Win32'">
-    <OutDir>.\libs\$(Configuration)\</OutDir>
+    <OutDir>.\libs\$(Configuration)_x86\</OutDir>
     <IntDir>
     </IntDir>
   </PropertyGroup>
   <PropertyGroup Condition="'$(Configuration)|$(Platform)'=='Release|Win32'">
-    <OutDir>.\libs\$(Configuration)\</OutDir>
+    <OutDir>.\libs\$(Configuration)_x86\</OutDir>
     <IntDir>
     </IntDir>
   </PropertyGroup>
@@ -115,7 +115,7 @@
     <ClCompile>
       <WarningLevel>Level3</WarningLevel>
       <Optimization>Disabled</Optimization>
-      <SDLCheck>true</SDLCheck>
+      <SDLCheck>false</SDLCheck>
       <PreprocessorDefinitions>ASMJIT_STATIC;_MBCS;%(PreprocessorDefinitions)</PreprocessorDefinitions>
     </ClCompile>
     <Link>
@@ -139,7 +139,7 @@
       <Optimization>MaxSpeed</Optimization>
       <FunctionLevelLinking>true</FunctionLevelLinking>
       <IntrinsicFunctions>true</IntrinsicFunctions>
-      <SDLCheck>true</SDLCheck>
+      <SDLCheck>false</SDLCheck>
       <PreprocessorDefinitions>ASMJIT_STATIC;_UNICODE;UNICODE;%(PreprocessorDefinitions)</PreprocessorDefinitions>
     </ClCompile>
     <Link>
@@ -154,7 +154,7 @@
       <Optimization>MaxSpeed</Optimization>
       <FunctionLevelLinking>true</FunctionLevelLinking>
       <IntrinsicFunctions>true</IntrinsicFunctions>
-      <SDLCheck>true</SDLCheck>
+      <SDLCheck>false</SDLCheck>
       <PreprocessorDefinitions>ASMJIT_STATIC;_UNICODE;UNICODE;%(PreprocessorDefinitions)</PreprocessorDefinitions>
     </ClCompile>
     <Link>

rpcs3/Emu/Cell/PPCDecoder.h

@@ -12,19 +12,19 @@ public:
 
 
 template<typename TO, uint from, uint to>
-static InstrList<1 << CodeField<from, to>::size, TO>* new_list(const CodeField<from, to>& func, InstrCaller<TO>* error_func = nullptr)
+static InstrList<(1 << (CodeField<from, to>::size)), TO>* new_list(const CodeField<from, to>& func, InstrCaller<TO>* error_func = nullptr)
 {
-	return new InstrList<1 << CodeField<from, to>::size, TO>(func, error_func);
+	return new InstrList<(1 << (CodeField<from, to>::size)), TO>(func, error_func);
 }
 
 template<int count, typename TO, uint from, uint to>
-static InstrList<1 << CodeField<from, to>::size, TO>* new_list(InstrList<count, TO>* parent, int opcode, const CodeField<from, to>& func, InstrCaller<TO>* error_func = nullptr)
+static InstrList<(1 << (CodeField<from, to>::size)), TO>* new_list(InstrList<count, TO>* parent, int opcode, const CodeField<from, to>& func, InstrCaller<TO>* error_func = nullptr)
 {
-	return connect_list(parent, new InstrList<1 << CodeField<from, to>::size, TO>(func, error_func), opcode);
+	return connect_list(parent, new InstrList<(1 << (CodeField<from, to>::size)), TO>(func, error_func), opcode);
 }
 
 template<int count, typename TO, uint from, uint to>
-static InstrList<1 << CodeField<from, to>::size, TO>* new_list(InstrList<count, TO>* parent, const CodeField<from, to>& func, InstrCaller<TO>* error_func = nullptr)
+static InstrList<(1 << (CodeField<from, to>::size)), TO>* new_list(InstrList<count, TO>* parent, const CodeField<from, to>& func, InstrCaller<TO>* error_func = nullptr)
 {
-	return connect_list(parent, new InstrList<1 << CodeField<from, to>::size, TO>(func, error_func));
+	return connect_list(parent, new InstrList<(1 << (CodeField<from, to>::size)), TO>(func, error_func));
 }

rpcs3/Emu/Cell/SPUInterpreter.h

@@ -32,121 +32,19 @@ private:
 	//0 - 10
 	void STOP(u32 code)
 	{
-		CPU.SetExitStatus(code); // exit code (not status)
+		CPU.DoStop(code);
-
-		switch (code)
-		{
-		case 0x110: /* ===== sys_spu_thread_receive_event ===== */
-			{
-				u32 spuq = 0;
-				if (!CPU.SPU.Out_MBox.Pop(spuq))
-				{
-					ConLog.Error("sys_spu_thread_receive_event: cannot read Out_MBox");
-					CPU.SPU.In_MBox.PushUncond(CELL_EINVAL); // ???
-					return;
-				}
-
-				if (CPU.SPU.In_MBox.GetCount())
-				{
-					ConLog.Error("sys_spu_thread_receive_event(spuq=0x%x): In_MBox is not empty", spuq);
-					CPU.SPU.In_MBox.PushUncond(CELL_EBUSY); // ???
-					return;
-				}
-
-				if (Ini.HLELogging.GetValue())
-				{
-					ConLog.Write("sys_spu_thread_receive_event(spuq=0x%x)", spuq);
-				}
-
-				EventQueue* eq;
-				if (!CPU.SPUQs.GetEventQueue(FIX_SPUQ(spuq), eq))
-				{
-					CPU.SPU.In_MBox.PushUncond(CELL_EINVAL); // TODO: check error value
-					return;
-				}
-
-				u32 tid = GetCurrentSPUThread().GetId();
-
-				eq->sq.push(tid); // add thread to sleep queue
-
-				while (true)
-				{
-					switch (eq->owner.trylock(tid))
-					{
-					case SMR_OK:
-						if (!eq->events.count())
-						{
-							eq->owner.unlock(tid);
-							break;
-						}
-						else
-						{
-							u32 next = (eq->protocol == SYS_SYNC_FIFO) ? eq->sq.pop() : eq->sq.pop_prio();
-							if (next != tid)
-							{
-								eq->owner.unlock(tid, next);
-								break;
-							}
-						}
-					case SMR_SIGNAL:
-						{
-							sys_event_data event;
-							eq->events.pop(event);
-							eq->owner.unlock(tid);
-							CPU.SPU.In_MBox.PushUncond(CELL_OK);
-							CPU.SPU.In_MBox.PushUncond(event.data1);
-							CPU.SPU.In_MBox.PushUncond(event.data2);
-							CPU.SPU.In_MBox.PushUncond(event.data3);
-							return;
-						}
-					case SMR_FAILED: break;
-					default: eq->sq.invalidate(tid); CPU.SPU.In_MBox.PushUncond(CELL_ECANCELED); return;
-					}
-
-					Sleep(1);
-					if (Emu.IsStopped())
-					{
-						ConLog.Warning("sys_spu_thread_receive_event(spuq=0x%x) aborted", spuq);
-						eq->sq.invalidate(tid);
-						return;
-					}
-				}
-			}
-			break;
-		case 0x102:
-			if (!CPU.SPU.Out_MBox.GetCount())
-			{
-				ConLog.Error("sys_spu_thread_exit (no status, code 0x102)");
-			}
-			else if (Ini.HLELogging.GetValue())
-			{
-				// the real exit status
-				ConLog.Write("sys_spu_thread_exit (status=0x%x)", CPU.SPU.Out_MBox.GetValue());
-			}
-			CPU.Stop();
-			break;
-		default:
-			if (!CPU.SPU.Out_MBox.GetCount())
-			{
-				ConLog.Error("Unknown STOP code: 0x%x (no message)", code);
-			}
-			else
-			{
-				ConLog.Error("Unknown STOP code: 0x%x (message=0x%x)", code, CPU.SPU.Out_MBox.GetValue());
-			}
-			CPU.Stop();
-			break;
-		}
 	}
 	void LNOP()
 	{
 	}
 	void SYNC(u32 Cbit)
 	{
+		// This instruction must be used following a store instruction that modifies the instruction stream.
 		_mm_mfence();
 	}
 	void DSYNC()
 	{
+		// This instruction forces all earlier load, store, and channel instructions to complete before proceeding.
 		_mm_mfence();
 	}
 	void MFSPR(u32 rt, u32 sa)
@@ -389,6 +287,7 @@ private:
 	}
 	void STOPD(u32 rc, u32 ra, u32 rb)
 	{
+		UNIMPLEMENTED();
 		Emu.Pause();
 	}
 	void STQX(u32 rt, u32 ra, u32 rb)

rpcs3/Emu/Cell/SPURecompilerCore.cpp

@@ -7,7 +7,7 @@ static const SPUImmTable g_spu_imm;
 
 SPURecompilerCore::SPURecompilerCore(SPUThread& cpu)
 : m_enc(new SPURecompiler(cpu, *this))
-, m_inter(new SPUInterpreter(cpu))
+, inter(new SPUInterpreter(cpu))
 , CPU(cpu)
 , compiler(&runtime)
 {
@@ -17,17 +17,17 @@ SPURecompilerCore::SPURecompilerCore(SPUThread& cpu)
 SPURecompilerCore::~SPURecompilerCore()
 {
 	delete m_enc;
-	delete m_inter;
+	delete inter;
 }
 
 void SPURecompilerCore::Decode(const u32 code) // decode instruction and run with interpreter
 {
-	(*SPU_instr::rrr_list)(m_inter, code);
+	(*SPU_instr::rrr_list)(inter, code);
 }
 
 void SPURecompilerCore::Compile(u16 pos)
 {
-	compiler.addFunc(kFuncConvHost, FuncBuilder4<u16, void*, void*, void*, u16>());
+	compiler.addFunc(kFuncConvHost, FuncBuilder4<u32, void*, void*, void*, u32>());
 	entry[pos].host = pos;
 
 	GpVar cpu_var(compiler, kVarTypeIntPtr, "cpu");
@@ -45,15 +45,26 @@ void SPURecompilerCore::Compile(u16 pos)
 	compiler.alloc(imm_var);
 	m_enc->imm_var = &imm_var;
 
-	GpVar pos_var(compiler, kVarTypeUInt16, "pos");
+	GpVar pos_var(compiler, kVarTypeUInt32, "pos");
 	compiler.setArg(3, pos_var);
 	compiler.alloc(pos_var);
 
+	m_enc->pos_var = &pos_var;
+
+	compiler.xor_(pos_var, pos_var);
+
 	while (true)
 	{
 		const u32 opcode = Memory.Read32(CPU.dmac.ls_offset + pos * 4);
 		m_enc->do_finalize = false;
+		if (opcode)
+		{
 			(*SPU_instr::rrr_list)(m_enc, opcode); // compile single opcode
+		}
+		else
+		{
+			m_enc->do_finalize = true;
+		}
 		bool fin = m_enc->do_finalize;
 		entry[pos].valid = opcode;
 
@@ -63,7 +74,6 @@ void SPURecompilerCore::Compile(u16 pos)
 		entry[pos].host = entry[pos - 1].host;
 	}
 
-	compiler.xor_(pos_var, pos_var);
 	compiler.ret(pos_var);
 	compiler.endFunc();
 	entry[entry[pos].host].pointer = compiler.make();
@@ -74,6 +84,7 @@ u8 SPURecompilerCore::DecodeMemory(const u64 address)
 	const u64 m_offset = address - CPU.PC;
 	const u16 pos = (CPU.PC >> 2);
 
+	//ConLog.Write("DecodeMemory: pos=%d", pos);
 	u32* ls = (u32*)Memory.VirtualToRealAddr(m_offset);
 
 	if (!pos)
@@ -115,16 +126,16 @@ u8 SPURecompilerCore::DecodeMemory(const u64 address)
 		return 0;
 	}
 	// jump
-	typedef u16(*Func)(void* _cpu, void* _ls, const SPUImmTable* _imm, u16 _pos);
+	typedef u32(*Func)(void* _cpu, void* _ls, const SPUImmTable* _imm, u32 _pos);
 
 	Func func = asmjit_cast<Func>(entry[entry[pos].host].pointer);
 
 	void* cpu = (u8*)&CPU.GPR[0] - offsetof(SPUThread, GPR[0]); // ugly cpu base offset detection
 
-	u16 res = pos == entry[pos].host ? 0 : pos;
+	u16 res = (pos == entry[pos].host) ? 0 : pos;
-	res = func(cpu, ls, &g_spu_imm, res);
+	res = (u16)func(cpu, ls, &g_spu_imm, res);
 
-	ConLog.Write("func -> %d", res);
+	CPU.SetBranch((u64)res << 2);
 
 	return 0;
 	/*Decode(Memory.Read32(address));

rpcs3/Emu/Cell/SPUThread.cpp

@@ -3,6 +3,7 @@
 #include "Emu/Cell/SPUDecoder.h"
 #include "Emu/Cell/SPUInterpreter.h"
 #include "Emu/Cell/SPUDisAsm.h"
+#include "Emu/Cell/SPURecompiler.h"
 
 SPUThread& GetCurrentSPUThread()
 {
@@ -75,6 +76,8 @@ void SPUThread::DoRun()
 	break;
 
 	case 1:
+		m_dec = new SPURecompilerCore(*this);
+	break;
 	case 2:
 		m_dec = new SPUDecoder(*new SPUInterpreter(*this));
 	break;

rpcs3/Emu/Cell/SPUThread.h

@@ -213,20 +213,21 @@ public:
 
 union SPU_GPR_hdr
 {
+	u32 _u32[4];
+	float _f[4];
 	u128 _u128;
 	s128 _i128;
 	__m128 _m128;
 	__m128i _m128i;
 	u64 _u64[2];
 	s64 _i64[2];
-	u32 _u32[4];
 	s32 _i32[4];
 	u16 _u16[8];
 	s16 _i16[8];
 	u8  _u8[16];
 	s8  _i8[16];
 	double _d[2];
-	float _f[4];
+
 
 	SPU_GPR_hdr() {}
 
@@ -243,9 +244,9 @@ union SPU_GPR_hdr
 
 union SPU_SPR_hdr
 {
+	u32 _u32[4];
 	u128 _u128;
 	s128 _i128;
-	u32 _u32[4];
 
 	SPU_SPR_hdr() {}
 
@@ -299,19 +300,19 @@ public:
 #else
 		static const bool x86 = true;
 #endif
-
-	private:
 		union _CRT_ALIGN(8) {
 			struct {
 				volatile u32 m_index;
 				u32 m_value[max_count];
 			};
+			struct {
+				volatile u32 m_index2;
+				u16 m_val16[max_count * 2];
+			};
 			volatile u64 m_indval;
 		};
 		std::mutex m_lock;
 
-	public:
-
 		Channel()
 		{
 			Init();
@@ -586,7 +587,7 @@ public:
 			}
 		}
 
-		Sleep(1); // hack
+		//Sleep(1); // hack
 
 		switch(cmd & ~(MFC_BARRIER_MASK | MFC_FENCE_MASK | MFC_LIST_MASK | MFC_RESULT_MASK))
 		{
@@ -1125,6 +1126,115 @@ public:
 		if (Emu.IsStopped()) ConLog.Warning("%s(%s) aborted", __FUNCTION__, spu_ch_name[ch]);
 	}
 
+	void DoStop(u32 code)
+	{
+		SetExitStatus(code); // exit code (not status)
+
+		switch (code)
+		{
+		case 0x110: /* ===== sys_spu_thread_receive_event ===== */
+		{
+			u32 spuq = 0;
+			if (!SPU.Out_MBox.Pop(spuq))
+			{
+				ConLog.Error("sys_spu_thread_receive_event: cannot read Out_MBox");
+				SPU.In_MBox.PushUncond(CELL_EINVAL); // ???
+				return;
+			}
+
+			if (SPU.In_MBox.GetCount())
+			{
+				ConLog.Error("sys_spu_thread_receive_event(spuq=0x%x): In_MBox is not empty", spuq);
+				SPU.In_MBox.PushUncond(CELL_EBUSY); // ???
+				return;
+			}
+
+			if (Ini.HLELogging.GetValue())
+			{
+				ConLog.Write("sys_spu_thread_receive_event(spuq=0x%x)", spuq);
+			}
+
+			EventQueue* eq;
+			if (!SPUQs.GetEventQueue(FIX_SPUQ(spuq), eq))
+			{
+				SPU.In_MBox.PushUncond(CELL_EINVAL); // TODO: check error value
+				return;
+			}
+
+			u32 tid = GetId();
+
+			eq->sq.push(tid); // add thread to sleep queue
+
+			while (true)
+			{
+				switch (eq->owner.trylock(tid))
+				{
+				case SMR_OK:
+					if (!eq->events.count())
+					{
+						eq->owner.unlock(tid);
+						break;
+					}
+					else
+					{
+						u32 next = (eq->protocol == SYS_SYNC_FIFO) ? eq->sq.pop() : eq->sq.pop_prio();
+						if (next != tid)
+						{
+							eq->owner.unlock(tid, next);
+							break;
+						}
+					}
+				case SMR_SIGNAL:
+				{
+					sys_event_data event;
+					eq->events.pop(event);
+					eq->owner.unlock(tid);
+					SPU.In_MBox.PushUncond(CELL_OK);
+					SPU.In_MBox.PushUncond(event.data1);
+					SPU.In_MBox.PushUncond(event.data2);
+					SPU.In_MBox.PushUncond(event.data3);
+					return;
+				}
+				case SMR_FAILED: break;
+				default: eq->sq.invalidate(tid); SPU.In_MBox.PushUncond(CELL_ECANCELED); return;
+				}
+
+				Sleep(1);
+				if (Emu.IsStopped())
+				{
+					ConLog.Warning("sys_spu_thread_receive_event(spuq=0x%x) aborted", spuq);
+					eq->sq.invalidate(tid);
+					return;
+				}
+			}
+		}
+			break;
+		case 0x102:
+			if (!SPU.Out_MBox.GetCount())
+			{
+				ConLog.Error("sys_spu_thread_exit (no status, code 0x102)");
+			}
+			else if (Ini.HLELogging.GetValue())
+			{
+				// the real exit status
+				ConLog.Write("sys_spu_thread_exit (status=0x%x)", SPU.Out_MBox.GetValue());
+			}
+			Stop();
+			break;
+		default:
+			if (!SPU.Out_MBox.GetCount())
+			{
+				ConLog.Error("Unknown STOP code: 0x%x (no message)", code);
+			}
+			else
+			{
+				ConLog.Error("Unknown STOP code: 0x%x (message=0x%x)", code, SPU.Out_MBox.GetValue());
+			}
+			Stop();
+			break;
+		}
+	}
+
 	bool IsGoodLSA(const u32 lsa) const { return Memory.IsGoodAddr(lsa + m_offset) && lsa < 0x40000; }
 	virtual u8   ReadLS8  (const u32 lsa) const { return Memory.Read8  (lsa + m_offset); } // m_offset & 0x3fffc ?????
 	virtual u16  ReadLS16 (const u32 lsa) const { return Memory.Read16 (lsa + m_offset); }

rpcs3/rpcs3.vcxproj

@@ -393,6 +393,7 @@
     <ClInclude Include="Emu\Cell\SPUDisAsm.h" />
     <ClInclude Include="Emu\Cell\SPUInterpreter.h" />
     <ClInclude Include="Emu\Cell\SPUOpcodes.h" />
+    <ClInclude Include="Emu\Cell\SPURecompiler.h" />
     <ClInclude Include="Emu\Cell\SPURSManager.h" />
     <ClInclude Include="Emu\Cell\SPUThread.h" />
     <ClInclude Include="Emu\DbgConsole.h" />

rpcs3/rpcs3.vcxproj.filters

@@ -702,5 +702,8 @@
     <ClInclude Include="..\Utilities\StrFmt.h">
       <Filter>Utilities</Filter>
     </ClInclude>
+    <ClInclude Include="Emu\Cell\SPURecompiler.h">
+      <Filter>Include</Filter>
+    </ClInclude>
   </ItemGroup>
 </Project>