Jit64: implement dcbf + dcbi

Source/Core/Core/HW/DSP.cpp

@@ -145,12 +145,12 @@ struct ARAMInfo
 
 // STATE_TO_SAVE
 static ARAMInfo g_ARAM;
-static UDSPControl g_dspState;
 static AudioDMA g_audioDMA;
 static ARAM_DMA g_arDMA;
 static u32 last_mmaddr;
 static u32 last_aram_dma_count;
 static bool instant_dma;
+UDSPControl g_dspState;
 
 union ARAM_Info
 {
@@ -216,6 +216,22 @@ void EnableInstantDMA()
 	instant_dma = true;
 }
 
+void FlushInstantDMA(u32 address)
+{
+	u64 dma_in_progress = DSP::DMAInProgress();
+	if (dma_in_progress != 0)
+	{
+		u32 start_addr = (dma_in_progress >> 32) & Memory::RAM_MASK;
+		u32 end_addr = (dma_in_progress & Memory::RAM_MASK) & 0xffffffff;
+		u32 invalidated_addr = (address & Memory::RAM_MASK) & ~0x1f;
+
+		if (invalidated_addr >= start_addr && invalidated_addr <= end_addr)
+		{
+			DSP::EnableInstantDMA();
+		}
+	}
+}
+
 DSPEmulator *GetDSPEmulator()
 {
 	return dsp_emulator;

Source/Core/Core/HW/DSP.h

@@ -56,6 +56,8 @@ union UDSPControl
 	UDSPControl(u16 _Hex = 0) : Hex(_Hex) {}
 };
 
+extern UDSPControl g_dspState;
+
 void Init(bool hle);
 void Shutdown();
 
@@ -78,5 +80,6 @@ void UpdateAudioDMA();
 void UpdateDSPSlice(int cycles);
 u64 DMAInProgress();
 void EnableInstantDMA();
+void FlushInstantDMA(u32 address);
 
 }// end of namespace DSP

Source/Core/Core/PowerPC/Interpreter/Interpreter_LoadStore.cpp

@@ -340,18 +340,7 @@ void Interpreter::dcbi(UGeckoInstruction _inst)
 	// The following detects a situation where the game is writing to the dcache at the address being DMA'd. As we do not
 	// have dcache emulation, invalid data is being DMA'd causing audio glitches. The following code detects this and
 	// enables the DMA to complete instantly before the invalid data is written. Resident Evil 2 & 3 trigger this.
-	u64 dma_in_progress = DSP::DMAInProgress();
+	DSP::FlushInstantDMA(address);
-	if (dma_in_progress != 0)
-	{
-		u32 start_addr = (dma_in_progress >> 32) & Memory::RAM_MASK;
-		u32 end_addr = (dma_in_progress & Memory::RAM_MASK) & 0xffffffff;
-		u32 invalidated_addr = (address & Memory::RAM_MASK) & ~0x1f;
-
-		if (invalidated_addr >= start_addr && invalidated_addr <= end_addr)
-		{
-			DSP::EnableInstantDMA();
-		}
-	}
 }
 
 void Interpreter::dcbst(UGeckoInstruction _inst)

Source/Core/Core/PowerPC/Jit64/Jit.h

@@ -255,4 +255,6 @@ public:
 
 	void lmw(UGeckoInstruction inst);
 	void stmw(UGeckoInstruction inst);
+
+	void dcbx(UGeckoInstruction inst);
 };

Source/Core/Core/PowerPC/Jit64/Jit64_Tables.cpp

@@ -214,11 +214,11 @@ static GekkoOPTemplate table31[] =
 	{824, &Jit64::srawix},                 // srawix
 	{24,  &Jit64::slwx},                   // slwx
 
-	{54,   &Jit64::FallBackToInterpreter}, // dcbst
+	{54,   &Jit64::dcbx},                  // dcbst
-	{86,   &Jit64::FallBackToInterpreter}, // dcbf
+	{86,   &Jit64::dcbx},                  // dcbf
-	{246,  &Jit64::dcbt },                 // dcbtst
+	{246,  &Jit64::dcbt},                  // dcbtst
-	{278,  &Jit64::dcbt },                 // dcbt
+	{278,  &Jit64::dcbt},                  // dcbt
-	{470,  &Jit64::FallBackToInterpreter}, // dcbi
+	{470,  &Jit64::dcbx},                  // dcbi
 	{758,  &Jit64::DoNothing},             // dcba
 	{1014, &Jit64::dcbz},                  // dcbz
 

Source/Core/Core/PowerPC/Jit64/Jit_LoadStore.cpp

@@ -7,6 +7,8 @@
 
 #include "Common/CommonTypes.h"
 
+#include "Core/HW/DSP.h"
+#include "Core/PowerPC/JitInterface.h"
 #include "Core/PowerPC/Jit64/Jit.h"
 #include "Core/PowerPC/Jit64/JitAsm.h"
 #include "Core/PowerPC/Jit64/JitRegCache.h"
@@ -290,6 +292,68 @@ void Jit64::lXXx(UGeckoInstruction inst)
 	gpr.UnlockAllX();
 }
 
+void Jit64::dcbx(UGeckoInstruction inst)
+{
+	INSTRUCTION_START
+	JITDISABLE(bJITLoadStoreOff);
+
+	X64Reg addr = RSCRATCH;
+	X64Reg value = RSCRATCH2;
+	X64Reg tmp = ECX;
+
+	PUSH(tmp);
+
+	MOV(32, R(addr), gpr.R(inst.RB));
+	if (inst.RA)
+	{
+		ADD(32, R(addr), gpr.R(inst.RA));
+	}
+
+	MOV(32, R(value), R(addr));
+	SHL(32, R(value), Imm8(3));
+	SHR(32, R(value), Imm8(13));
+	MOV(64, R(tmp), ImmPtr(jit->GetBlockCache()->GetBlockBitSet()));
+	MOV(32, R(value), MComplex(tmp, value, SCALE_4, 0));
+
+	MOV(32, R(tmp), R(addr));
+	SHR(32, R(tmp), Imm8(5));
+	SHR(32, R(value), R(tmp));
+	TEST(32, R(value), Imm32(1));
+
+	FixupBranch c = J_CC(CC_NZ, true);
+	SwitchToFarCode();
+	SetJumpTarget(c);
+	BitSet32 registersInUse = CallerSavedRegistersInUse();
+	ABI_PushRegistersAndAdjustStack(registersInUse, 0);
+	MOV(32, R(ABI_PARAM1), R(addr));
+	MOV(32, R(ABI_PARAM2), Imm32(32));
+	XOR(32, R(ABI_PARAM3), R(ABI_PARAM3));
+	ABI_CallFunction((void*)JitInterface::InvalidateICache);
+	ABI_PopRegistersAndAdjustStack(registersInUse, 0);
+	c = J(true);
+	SwitchToNearCode();
+	SetJumpTarget(c);
+
+	// dcbi
+	if (inst.SUBOP10 == 470)
+	{
+		MOV(16, R(tmp), M(&DSP::g_dspState));
+		TEST(16, R(tmp), Imm16(1 << 9));
+		c = J_CC(CC_NZ, true);
+		SwitchToFarCode();
+		SetJumpTarget(c);
+		ABI_PushRegistersAndAdjustStack(registersInUse, 0);
+		MOV(32, R(ABI_PARAM1), R(addr));
+		ABI_CallFunction((void*)DSP::FlushInstantDMA);
+		ABI_PopRegistersAndAdjustStack(registersInUse, 0);
+		c = J(true);
+		SwitchToNearCode();
+		SetJumpTarget(c);
+	}
+
+	POP(tmp);
+}
+
 void Jit64::dcbt(UGeckoInstruction inst)
 {
 	INSTRUCTION_START

Source/Core/Core/PowerPC/JitCommon/JitCache.h

@@ -60,6 +60,7 @@ typedef void (*CompiledCode)();
 // implementation of std::bitset is slow.
 class ValidBlockBitSet final
 {
+public:
 	enum
 	{
 		VALID_BLOCK_MASK_SIZE = 0x20000000 / 32,
@@ -67,7 +68,6 @@ class ValidBlockBitSet final
 	};
 	std::unique_ptr<u32[]> m_valid_block;
 
-public:
 	ValidBlockBitSet()
 	{
 		m_valid_block.reset(new u32[VALID_BLOCK_ALLOC_ELEMENTS]);
@@ -157,6 +157,11 @@ public:
 
 	// DOES NOT WORK CORRECTLY WITH INLINING
 	void InvalidateICache(u32 address, const u32 length, bool forced);
+
+	u32* GetBlockBitSet() const
+	{
+		return valid_block.m_valid_block.get();
+	}
 };
 
 // x86 BlockCache