Fix all the current issues with the AArch64 register cache.

Fixes all the current issues I've been experiencing.
Scaled back the register cache idea for now so I can actually work on some real instructions.

Tested this work with unit tests so I know it works.
Unit tests are pretty great things.

Source/Core/Core/PowerPC/JitArm64/JitArm64_RegCache.cpp

@@ -13,8 +13,12 @@ void Arm64RegCache::Init(ARM64XEmitter *emitter)
 	GetAllocationOrder();
 }
 
-ARM64Reg Arm64RegCache::GetReg(void)
+ARM64Reg Arm64RegCache::GetReg()
 {
+	// If we have no registers left, dump the most stale register first
+	if (!GetUnlockedRegisterCount())
+		FlushMostStaleRegister();
+
 	for (auto& it : m_host_registers)
 	{
 		if (!it.IsLocked())
@@ -29,6 +33,15 @@ ARM64Reg Arm64RegCache::GetReg(void)
 	return INVALID_REG;
 }
 
+u32 Arm64RegCache::GetUnlockedRegisterCount()
+{
+	u32 unlocked_registers = 0;
+	for (auto& it : m_host_registers)
+		if (!it.IsLocked())
+			++unlocked_registers;
+	return unlocked_registers;
+}
+
 void Arm64RegCache::LockRegister(ARM64Reg host_reg)
 {
 	auto reg = std::find(m_host_registers.begin(), m_host_registers.end(), host_reg);
@@ -101,6 +114,81 @@ bool Arm64GPRCache::IsCalleeSaved(ARM64Reg reg)
 	return std::find(callee_regs.begin(), callee_regs.end(), EncodeRegTo64(reg)) != callee_regs.end();
 }
 
+void Arm64GPRCache::FlushRegister(u32 preg)
+{
+	u32 base_reg = preg;
+	OpArg& reg = m_guest_registers[preg];
+	if (reg.GetType() == REG_REG)
+	{
+		ARM64Reg host_reg = reg.GetReg();
+
+		m_emit->STR(INDEX_UNSIGNED, host_reg, X29, PPCSTATE_OFF(gpr[preg]));
+		Unlock(host_reg);
+
+		reg.Flush();
+	}
+	else if (reg.GetType() == REG_IMM)
+	{
+		ARM64Reg host_reg = GetReg();
+
+		m_emit->MOVI2R(host_reg, reg.GetImm());
+		m_emit->STR(INDEX_UNSIGNED, host_reg, X29, PPCSTATE_OFF(gpr[preg]));
+
+		Unlock(host_reg);
+
+		reg.Flush();
+	}
+	else if (reg.GetType() == REG_AWAY)
+	{
+		u32 next_reg = 0;
+		if (reg.GetAwayLocation() == REG_LOW)
+			next_reg = base_reg + 1;
+		else
+			next_reg = base_reg - 1;
+		OpArg& reg2 = m_guest_registers[next_reg];
+		ARM64Reg host_reg = reg.GetAwayReg();
+		ARM64Reg host_reg_1 = reg.GetReg();
+		ARM64Reg host_reg_2 = reg2.GetReg();
+		// Flush if either of these shared registers are used.
+		if (host_reg_1 == INVALID_REG)
+		{
+			// We never loaded this register
+			// We've got to test the state of our shared register
+			// Currently it is always reg+1
+			if (host_reg_2 == INVALID_REG)
+			{
+				// We didn't load either of these registers
+				// This can happen in cases where we had to flush register state
+				// or if we hit an interpreted instruction before we could use it
+				// Dump the whole thing in one go and flush both registers
+
+				// 64bit host register will store 2 32bit store registers in one go
+				if (reg.GetAwayLocation() == REG_LOW)
+					m_emit->STR(INDEX_UNSIGNED, host_reg, X29, PPCSTATE_OFF(gpr[base_reg]));
+				else
+					m_emit->STR(INDEX_UNSIGNED, host_reg, X29, PPCSTATE_OFF(gpr[next_reg]));
+			}
+			else
+			{
+				// Alright, bottom register isn't used, but top one is
+				// Only store the top one
+				m_emit->STR(INDEX_UNSIGNED, host_reg_2, X29, PPCSTATE_OFF(gpr[next_reg]));
+				Unlock(host_reg_2);
+			}
+		}
+		else
+		{
+			m_emit->STR(INDEX_UNSIGNED, host_reg_1, X29, PPCSTATE_OFF(gpr[base_reg]));
+			Unlock(host_reg_1);
+		}
+		// Flush both registers
+		reg.Flush();
+		reg2.Flush();
+		Unlock(DecodeReg(host_reg));
+	}
+
+}
+
 void Arm64GPRCache::Flush(FlushMode mode, PPCAnalyst::CodeOp* op)
 {
 	for (int i = 0; i < 32; ++i)
@@ -124,26 +212,12 @@ void Arm64GPRCache::Flush(FlushMode mode, PPCAnalyst::CodeOp* op)
 			// Has to be flushed if it isn't in a callee saved register
 			ARM64Reg host_reg = m_guest_registers[i].GetReg();
 			if (flush || !IsCalleeSaved(host_reg))
-			{
+				FlushRegister(i);
-				m_emit->STR(INDEX_UNSIGNED, host_reg, X29, PPCSTATE_OFF(gpr[i]));
-				Unlock(host_reg);
-
-				m_guest_registers[i].Flush();
-			}
 		}
 		else if (m_guest_registers[i].GetType() == REG_IMM)
 		{
 			if (flush)
-			{
+				FlushRegister(i);
-				ARM64Reg host_reg = GetReg();
-
-				m_emit->MOVI2R(host_reg, m_guest_registers[i].GetImm());
-				m_emit->STR(INDEX_UNSIGNED, host_reg, X29, PPCSTATE_OFF(gpr[i]));
-
-				Unlock(host_reg);
-
-				m_guest_registers[i].Flush();
-			}
 		}
 		else if (m_guest_registers[i].GetType() == REG_AWAY)
 		{
@@ -173,39 +247,7 @@ void Arm64GPRCache::Flush(FlushMode mode, PPCAnalyst::CodeOp* op)
 			    !IsCalleeSaved(host_reg_1) ||
 			    !IsCalleeSaved(host_reg_2))
 			{
-
+				FlushRegister(i); // Will flush both pairs of registers
-				if (host_reg_1 == INVALID_REG)
-				{
-					// We never loaded this register
-					// We've got to test the state of our shared register
-					// Currently it is always reg+1
-					if (host_reg_2 == INVALID_REG)
-					{
-						// We didn't load either of these registers
-						// This can happen in cases where we had to flush register state
-						// or if we hit an interpreted instruction before we could use it
-						// Dump the whole thing in one go and flush both registers
-
-						// 64bit host register will store 2 32bit store registers in one go
-						m_emit->STR(INDEX_UNSIGNED, host_reg, X29, PPCSTATE_OFF(gpr[i]));
-					}
-					else
-					{
-						// Alright, bottom register isn't used, but top one is
-						// Only store the top one
-						m_emit->STR(INDEX_UNSIGNED, host_reg_2, X29, PPCSTATE_OFF(gpr[i + 1]));
-						Unlock(host_reg_2);
-					}
-				}
-				else
-				{
-					m_emit->STR(INDEX_UNSIGNED, host_reg_1, X29, PPCSTATE_OFF(gpr[i]));
-					Unlock(host_reg_1);
-				}
-				// Flush both registers
-				m_guest_registers[i].Flush();
-				m_guest_registers[i + 1].Flush();
-				Unlock(DecodeReg(host_reg));
 			}
 			// Skip the next register since we've handled it here
 			++i;
@@ -216,6 +258,9 @@ void Arm64GPRCache::Flush(FlushMode mode, PPCAnalyst::CodeOp* op)
 ARM64Reg Arm64GPRCache::R(u32 preg)
 {
 	OpArg& reg = m_guest_registers[preg];
+	IncrementAllUsed();
+	reg.ResetLastUsed();
+
 	switch (reg.GetType())
 	{
 	case REG_REG: // already in a reg
@@ -225,6 +270,8 @@ ARM64Reg Arm64GPRCache::R(u32 preg)
 	{
 		ARM64Reg host_reg = GetReg();
 		m_emit->MOVI2R(host_reg, reg.GetImm());
+		reg.LoadToReg(host_reg);
+		return host_reg;
 	}
 	break;
 	case REG_AWAY: // Register is away in a shared register
@@ -232,22 +279,57 @@ ARM64Reg Arm64GPRCache::R(u32 preg)
 		// Let's do the voodoo that we dodo
 		if (reg.GetReg() == INVALID_REG)
 		{
-			// Alright, we need to move to a valid location
-			ARM64Reg host_reg = GetReg();
-			reg.LoadAwayToReg(host_reg);
-
 			// Alright, we need to extract from our away register
 			// To our new 32bit register
 			if (reg.GetAwayLocation() == REG_LOW)
 			{
-				// We are in the low bits
+				OpArg& upper_reg = m_guest_registers[preg + 1];
-				// Just move it over to the low bits of the new register
+				if (upper_reg.GetType() == REG_REG)
-				m_emit->UBFM(EncodeRegTo64(host_reg), reg.GetAwayReg(), 0, 31);
+				{
+					// If the upper reg is already moved away, just claim this one as ours now
+					ARM64Reg host_reg = reg.GetAwayReg();
+					reg.LoadToReg(DecodeReg(host_reg));
+					return host_reg;
+				}
+				else
+				{
+					// Top register is still loaded
+					// Make sure to move to a new register
+					ARM64Reg host_reg = GetReg();
+					ARM64Reg current_reg = reg.GetAwayReg();
+					reg.LoadToReg(host_reg);
+
+					// We are in the low bits
+					// Just move it over to the low bits of the new register
+					m_emit->UBFM(EncodeRegTo64(host_reg), current_reg, 0, 31);
+					return host_reg;
+				}
 			}
 			else
 			{
-				// We are in the high bits
+				OpArg& lower_reg = m_guest_registers[preg - 1];
-				m_emit->UBFM(EncodeRegTo64(host_reg), reg.GetAwayReg(), 32, 63);
+				if (lower_reg.GetType() == REG_REG)
+				{
+					// If the lower register is moved away, claim this one as ours
+					ARM64Reg host_reg = reg.GetAwayReg();
+					reg.LoadToReg(DecodeReg(host_reg));
+
+					// Make sure to move our register from the high bits to the low bits
+					m_emit->UBFM(EncodeRegTo64(host_reg), host_reg, 32, 63);
+					return host_reg;
+				}
+				else
+				{
+					// Load this register in to the new low bits
+					// We are no longer away
+					ARM64Reg host_reg = GetReg();
+					ARM64Reg current_reg = reg.GetAwayReg();
+					reg.LoadToReg(host_reg);
+
+					// We are in the high bits
+					m_emit->UBFM(EncodeRegTo64(host_reg), current_reg, 32, 63);
+					return host_reg;
+				}
 			}
 		}
 		else
@@ -259,10 +341,8 @@ ARM64Reg Arm64GPRCache::R(u32 preg)
 	break;
 	case REG_NOTLOADED: // Register isn't loaded at /all/
 	{
-		// This is kind of annoying, we shouldn't have gotten here
+		// This is a bit annoying. We try to keep these preloaded as much as possible
-		// This can happen with instructions that use multiple registers(eg lmw)
+		// This can also happen on cases where PPCAnalyst isn't feeing us proper register usage statistics
-		// The PPCAnalyst needs to be modified to handle these cases
-		_dbg_assert_msg_(DYNA_REC, false, "Hit REG_NOTLOADED type oparg. Fix the PPCAnalyst");
 		ARM64Reg host_reg = GetReg();
 		reg.LoadToReg(host_reg);
 		m_emit->LDR(INDEX_UNSIGNED, host_reg, X29, PPCSTATE_OFF(gpr[preg]));
@@ -277,7 +357,7 @@ ARM64Reg Arm64GPRCache::R(u32 preg)
 	return INVALID_REG;
 }
 
-void Arm64GPRCache::GetAllocationOrder(void)
+void Arm64GPRCache::GetAllocationOrder()
 {
 	// Callee saved registers first in hopes that we will keep everything stored there first
 	const std::vector<ARM64Reg> allocation_order =
@@ -292,6 +372,24 @@ void Arm64GPRCache::GetAllocationOrder(void)
 		m_host_registers.push_back(HostReg(reg));
 }
 
+void Arm64GPRCache::FlushMostStaleRegister()
+{
+	u32 most_stale_preg = 0;
+	u32 most_stale_amount = 0;
+	for (u32 i = 0; i < 32; ++i)
+	{
+		u32 last_used = m_guest_registers[i].GetLastUsed();
+		if (last_used > most_stale_amount &&
+		    m_guest_registers[i].GetType() != REG_IMM &&
+		    m_guest_registers[i].GetType() != REG_NOTLOADED)
+		{
+			most_stale_preg = i;
+			most_stale_amount = last_used;
+		}
+	}
+	FlushRegister(most_stale_preg);
+}
+
 // FPR Cache
 void Arm64FPRCache::Flush(FlushMode mode, PPCAnalyst::CodeOp* op)
 {
@@ -303,7 +401,7 @@ ARM64Reg Arm64FPRCache::R(u32 preg)
 	// XXX: return a host reg holding a guest register
 }
 
-void Arm64FPRCache::GetAllocationOrder(void)
+void Arm64FPRCache::GetAllocationOrder()
 {
 	const std::vector<ARM64Reg> allocation_order =
 	{
@@ -317,3 +415,8 @@ void Arm64FPRCache::GetAllocationOrder(void)
 		m_host_registers.push_back(HostReg(reg));
 }
 
+void Arm64FPRCache::FlushMostStaleRegister()
+{
+	// XXX: Flush a register
+}
+

Source/Core/Core/PowerPC/JitArm64/JitArm64_RegCache.h

@@ -42,10 +42,9 @@ class OpArg
 {
 public:
 	OpArg()
+		: m_type(REG_NOTLOADED), m_reg(INVALID_REG),
+		  m_value(0), m_last_used(0)
 	{
-		m_type = REG_NOTLOADED;
-		m_reg = INVALID_REG;
-		m_value = 0;
 	}
 
 	RegType GetType()
@@ -73,30 +72,40 @@ public:
 	{
 		m_type = REG_REG;
 		m_reg = reg;
+
+		m_away_reg = INVALID_REG;
 	}
 	void LoadToAway(ARM64Reg reg, RegLocation location)
 	{
 		m_type = REG_AWAY;
-		m_reg = INVALID_REG;
 		m_away_reg = reg;
 		m_away_location = location;
-	}
+
-	void LoadAwayToReg(ARM64Reg reg)
+		m_reg = INVALID_REG;
-	{
-		// We are still an away type
-		// We just are also in another register
-		m_reg = reg;
 	}
 	void LoadToImm(u32 imm)
 	{
 		m_type = REG_IMM;
 		m_value = imm;
+
+		m_reg = INVALID_REG;
+		m_away_reg = INVALID_REG;
 	}
 	void Flush()
 	{
+		// Invalidate any previous information
 		m_type = REG_NOTLOADED;
+		m_reg = INVALID_REG;
+		m_away_reg = INVALID_REG;
+
+		// Arbitrarily large value that won't roll over on a lot of increments
+		m_last_used = 0xFFFF;
 	}
 
+	u32 GetLastUsed() { return m_last_used; }
+	void ResetLastUsed() { m_last_used = 0; }
+	void IncrementLastUsed() { ++m_last_used; }
+
 private:
 	// For REG_REG
 	RegType m_type; // store type
@@ -110,6 +119,8 @@ private:
 
 	// For REG_IMM
 	u32 m_value; // IMM value
+
+	u32 m_last_used;
 };
 
 class HostReg
@@ -117,10 +128,10 @@ class HostReg
 public:
 	HostReg() : m_reg(INVALID_REG), m_locked(false) {}
 	HostReg(ARM64Reg reg) : m_reg(reg), m_locked(false) {}
-	bool IsLocked(void) { return m_locked; }
+	bool IsLocked() { return m_locked; }
-	void Lock(void) { m_locked = true; }
+	void Lock() { m_locked = true; }
-	void Unlock(void) { m_locked = false; }
+	void Unlock() { m_locked = false; }
-	ARM64Reg GetReg(void) { return m_reg; }
+	ARM64Reg GetReg() { return m_reg; }
 
 	bool operator==(const ARM64Reg& reg)
 	{
@@ -135,7 +146,7 @@ private:
 class Arm64RegCache
 {
 public:
-	Arm64RegCache(void) : m_emit(nullptr), m_reg_stats(nullptr) {};
+	Arm64RegCache() : m_emit(nullptr), m_reg_stats(nullptr) {};
 	virtual ~Arm64RegCache() {};
 
 	void Init(ARM64XEmitter *emitter);
@@ -151,7 +162,7 @@ public:
 
 	// Returns a temporary register for use
 	// Requires unlocking after done
-	ARM64Reg GetReg(void);
+	ARM64Reg GetReg();
 
 	// Locks a register so a cache cannot use it
 	// Useful for function calls
@@ -177,7 +188,10 @@ public:
 
 protected:
 	// Get the order of the host registers
-	virtual void GetAllocationOrder(void) = 0;
+	virtual void GetAllocationOrder() = 0;
+
+	// Flushes the most stale register
+	virtual void FlushMostStaleRegister() = 0;
 
 	// Lock a register
 	void LockRegister(ARM64Reg host_reg);
@@ -185,6 +199,9 @@ protected:
 	// Unlock a register
 	void UnlockRegister(ARM64Reg host_reg);
 
+	// Get available host registers
+	u32 GetUnlockedRegisterCount();
+
 	// Code emitter
 	ARM64XEmitter *m_emit;
 
@@ -220,7 +237,10 @@ public:
 
 protected:
 	// Get the order of the host registers
-	void GetAllocationOrder(void);
+	void GetAllocationOrder();
+
+	// Flushes the most stale register
+	void FlushMostStaleRegister();
 
 	// Our guest GPRs
 	// PowerPC has 32 GPRs
@@ -228,6 +248,14 @@ protected:
 
 private:
 	bool IsCalleeSaved(ARM64Reg reg);
+
+	void IncrementAllUsed()
+	{
+		for (auto& reg : m_guest_registers)
+			reg.IncrementLastUsed();
+	}
+
+	void FlushRegister(u32 preg);
 };
 
 class Arm64FPRCache : public Arm64RegCache
@@ -243,7 +271,10 @@ public:
 
 protected:
 	// Get the order of the host registers
-	void GetAllocationOrder(void);
+	void GetAllocationOrder();
+
+	// Flushes the most stale register
+	void FlushMostStaleRegister();
 
 	// Our guest FPRs
 	// Gekko has 32 paired registers(32x2)