CPU/Recompiler: Flush caller-saved regs before calling functions

Should provide a very small performance boost.

src/core/cpu_recompiler_code_generator_aarch64.cpp

@@ -1288,6 +1288,7 @@ Value CodeGenerator::EmitLoadGuestMemory(const CodeBlockInstruction& cbi, const
   {
     // We need to use the full 64 bits here since we test the sign bit result.
     Value result = m_register_cache.AllocateScratch(RegSize_64);
+    m_register_cache.FlushCallerSavedGuestRegisters(true, true);
 
     // NOTE: This can leave junk in the upper bits
     switch (size)
@@ -1357,6 +1358,8 @@ Value CodeGenerator::EmitLoadGuestMemory(const CodeBlockInstruction& cbi, const
   else
   {
     Value result = m_register_cache.AllocateScratch(RegSize_32);
+    m_register_cache.FlushCallerSavedGuestRegisters(true, true);
+
     switch (size)
     {
       case RegSize_8:
@@ -1406,6 +1409,8 @@ void CodeGenerator::EmitStoreGuestMemory(const CodeBlockInstruction& cbi, const
   if (g_settings.cpu_recompiler_memory_exceptions)
   {
     Value result = m_register_cache.AllocateScratch(RegSize_32);
+    m_register_cache.FlushCallerSavedGuestRegisters(true, true);
+
     switch (value.size)
     {
       case RegSize_8:
@@ -1449,6 +1454,8 @@ void CodeGenerator::EmitStoreGuestMemory(const CodeBlockInstruction& cbi, const
   }
   else
   {
+    m_register_cache.FlushCallerSavedGuestRegisters(true, true);
+
     switch (value.size)
     {
       case RegSize_8:

src/core/cpu_recompiler_code_generator_x64.cpp

@@ -1744,6 +1744,7 @@ Value CodeGenerator::EmitLoadGuestMemory(const CodeBlockInstruction& cbi, const
   {
     // We need to use the full 64 bits here since we test the sign bit result.
     Value result = m_register_cache.AllocateScratch(RegSize_64);
+    m_register_cache.FlushCallerSavedGuestRegisters(true, true);
 
     // NOTE: This can leave junk in the upper bits
     switch (size)
@@ -1811,6 +1812,8 @@ Value CodeGenerator::EmitLoadGuestMemory(const CodeBlockInstruction& cbi, const
   else
   {
     Value result = m_register_cache.AllocateScratch(RegSize_32);
+    m_register_cache.FlushCallerSavedGuestRegisters(true, true);
+
     switch (size)
     {
       case RegSize_8:
@@ -1860,6 +1863,8 @@ void CodeGenerator::EmitStoreGuestMemory(const CodeBlockInstruction& cbi, const
   if (g_settings.cpu_recompiler_memory_exceptions)
   {
     Value result = m_register_cache.AllocateScratch(RegSize_32);
+    m_register_cache.FlushCallerSavedGuestRegisters(true, true);
+
     switch (value.size)
     {
       case RegSize_8:
@@ -1901,6 +1906,8 @@ void CodeGenerator::EmitStoreGuestMemory(const CodeBlockInstruction& cbi, const
   }
   else
   {
+    m_register_cache.FlushCallerSavedGuestRegisters(true, true);
+
     switch (value.size)
     {
       case RegSize_8:

src/core/cpu_recompiler_register_cache.cpp

@@ -739,6 +739,21 @@ void RegisterCache::FlushAllGuestRegisters(bool invalidate, bool clear_dirty)
     FlushGuestRegister(static_cast<Reg>(reg), invalidate, clear_dirty);
 }
 
+void RegisterCache::FlushCallerSavedGuestRegisters(bool invalidate, bool clear_dirty)
+{
+  for (u8 reg = 0; reg < static_cast<u8>(Reg::count); reg++)
+  {
+    const Value& gr = m_state.guest_reg_state[reg];
+    if (!gr.IsInHostRegister() ||
+        (m_state.host_reg_state[gr.GetHostRegister()] & HostRegState::CallerSaved) != HostRegState::CallerSaved)
+    {
+      continue;
+    }
+
+    FlushGuestRegister(static_cast<Reg>(reg), invalidate, clear_dirty);
+  }
+}
+
 bool RegisterCache::EvictOneGuestRegister()
 {
   if (m_state.guest_reg_order_count == 0)

src/core/cpu_recompiler_register_cache.h

@@ -316,6 +316,7 @@ public:
 
   void InvalidateAllNonDirtyGuestRegisters();
   void FlushAllGuestRegisters(bool invalidate, bool clear_dirty);
+  void FlushCallerSavedGuestRegisters(bool invalidate, bool clear_dirty);
   bool EvictOneGuestRegister();
 
   /// Temporarily prevents register allocation.