CPU/Recompiler: Use fastmem instead of global for RAM loads

src/core/bus.cpp

@@ -284,9 +284,9 @@ void UpdateFastmemViews(bool enabled, bool isolate_cache)
     CPU::g_state.fastmem_base = m_fastmem_base;
   }
 
-  auto MapRAM = [](u32 base_address) {
+  auto MapRAM = [](u32 base_address, bool writable) {
     u8* map_address = m_fastmem_base + base_address;
-    auto view = m_memory_arena.CreateView(MEMORY_ARENA_RAM_OFFSET, RAM_SIZE, true, false, map_address);
+    auto view = m_memory_arena.CreateView(MEMORY_ARENA_RAM_OFFSET, RAM_SIZE, writable, false, map_address);
     if (!view)
     {
       Log_ErrorPrintf("Failed to map RAM at fastmem area %p (offset 0x%08X)", map_address, RAM_SIZE);
@@ -313,23 +313,23 @@ void UpdateFastmemViews(bool enabled, bool isolate_cache)
   if (!isolate_cache)
   {
     // KUSEG - cached
-    MapRAM(0x00000000);
+    MapRAM(0x00000000, !isolate_cache);
-    //MapRAM(0x00200000);
+    //MapRAM(0x00200000, !isolate_cache);
-    //MapRAM(0x00400000);
+    //MapRAM(0x00400000, !isolate_cache);
-    //MapRAM(0x00600000);
+    //MapRAM(0x00600000, !isolate_cache);
 
     // KSEG0 - cached
-    MapRAM(0x80000000);
+    MapRAM(0x80000000, !isolate_cache);
-    //MapRAM(0x80200000);
+    //MapRAM(0x80200000, !isolate_cache);
-    //MapRAM(0x80400000);
+    //MapRAM(0x80400000, !isolate_cache);
-    //MapRAM(0x80600000);
+    //MapRAM(0x80600000, !isolate_cache);
   }
 
   // KSEG1 - uncached
-  MapRAM(0xA0000000);
+  MapRAM(0xA0000000, true);
-  //MapRAM(0xA0200000);
+  //MapRAM(0xA0200000, true);
-  //MapRAM(0xA0400000);
+  //MapRAM(0xA0400000, true);
-  //MapRAM(0xA0600000);
+  //MapRAM(0xA0600000, true);
 }
 
 bool CanUseFastmemForAddress(VirtualMemoryAddress address)

src/core/cpu_recompiler_code_generator.h

@@ -79,6 +79,7 @@ public:
   // Automatically generates an exception handler.
   Value EmitLoadGuestMemory(const CodeBlockInstruction& cbi, const Value& address, const SpeculativeValue& address_spec,
                             RegSize size);
+  void EmitLoadGuestRAMFastmem(const Value& address, RegSize size, Value& result);
   void EmitLoadGuestMemoryFastmem(const CodeBlockInstruction& cbi, const Value& address, RegSize size, Value& result);
   void EmitLoadGuestMemorySlowmem(const CodeBlockInstruction& cbi, const Value& address, RegSize size, Value& result,
                                   bool in_far_code);

src/core/cpu_recompiler_code_generator_aarch64.cpp

@@ -1281,6 +1281,39 @@ void CodeGenerator::EmitAddCPUStructField(u32 offset, const Value& value)
   }
 }
 
+void CodeGenerator::EmitLoadGuestRAMFastmem(const Value& address, RegSize size, Value& result)
+{
+  a64::MemOperand actual_address;
+  if (address.IsConstant())
+  {
+    m_emit->Mov(GetHostReg32(result.host_reg), address.constant_value);
+    actual_address = a64::MemOperand(GetFastmemBasePtrReg(), GetHostReg32(result.host_reg));
+  }
+  else
+  {
+    actual_address = a64::MemOperand(GetFastmemBasePtrReg(), GetHostReg32(address));
+  }
+
+  switch (size)
+  {
+  case RegSize_8:
+    m_emit->Ldrb(GetHostReg32(result.host_reg), actual_address);
+    break;
+
+  case RegSize_16:
+    m_emit->Ldrh(GetHostReg32(result.host_reg), actual_address);
+    break;
+
+  case RegSize_32:
+    m_emit->Ldr(GetHostReg32(result.host_reg), actual_address);
+    break;
+
+  default:
+    UnreachableCode();
+    break;
+  }
+}
+
 void CodeGenerator::EmitLoadGuestMemoryFastmem(const CodeBlockInstruction& cbi, const Value& address, RegSize size,
                                                Value& result)
 {

src/core/cpu_recompiler_code_generator_generic.cpp

@@ -1,8 +1,8 @@
+#include "common/log.h"
 #include "cpu_core.h"
 #include "cpu_core_private.h"
 #include "cpu_recompiler_code_generator.h"
 #include "settings.h"
-#include "common/log.h"
 Log_SetChannel(Recompiler::CodeGenerator);
 
 namespace CPU::Recompiler {
@@ -40,7 +40,12 @@ Value CodeGenerator::EmitLoadGuestMemory(const CodeBlockInstruction& cbi, const
     if (ptr)
     {
       Value result = m_register_cache.AllocateScratch(size);
-      EmitLoadGlobal(result.GetHostRegister(), size, ptr);
+
+      if (g_settings.IsUsingFastmem() && Bus::IsRAMAddress(static_cast<u32>(address.constant_value)))
+        EmitLoadGuestRAMFastmem(address, size, result);
+      else
+        EmitLoadGlobal(result.GetHostRegister(), size, ptr);
+
       m_delayed_cycles_add += read_ticks;
       return result;
     }

src/core/cpu_recompiler_code_generator_x64.cpp

@@ -1745,6 +1745,66 @@ void CodeGenerator::EmitAddCPUStructField(u32 offset, const Value& value)
   }
 }
 
+void CodeGenerator::EmitLoadGuestRAMFastmem(const Value& address, RegSize size, Value& result)
+{
+  // can't store displacements > 0x80000000 in-line
+  const Value* actual_address = &address;
+  if (address.IsConstant() && address.constant_value >= 0x80000000)
+  {
+    actual_address = &result;
+    m_emit->mov(GetHostReg32(result.host_reg), address.constant_value);
+  }
+
+  // TODO: movsx/zx inline here
+  switch (size)
+  {
+    case RegSize_8:
+    {
+      if (actual_address->IsConstant())
+      {
+        m_emit->mov(GetHostReg8(result.host_reg),
+                    m_emit->byte[GetFastmemBasePtrReg() + actual_address->constant_value]);
+      }
+      else
+      {
+        m_emit->mov(GetHostReg8(result.host_reg),
+                    m_emit->byte[GetFastmemBasePtrReg() + GetHostReg64(actual_address->host_reg)]);
+      }
+    }
+    break;
+
+    case RegSize_16:
+    {
+      if (actual_address->IsConstant())
+      {
+        m_emit->mov(GetHostReg16(result.host_reg),
+                    m_emit->word[GetFastmemBasePtrReg() + actual_address->constant_value]);
+      }
+      else
+      {
+        m_emit->mov(GetHostReg16(result.host_reg),
+                    m_emit->word[GetFastmemBasePtrReg() + GetHostReg64(actual_address->host_reg)]);
+      }
+    }
+    break;
+
+    case RegSize_32:
+    {
+      if (actual_address->IsConstant())
+      {
+        m_emit->mov(GetHostReg32(result.host_reg),
+                    m_emit->dword[GetFastmemBasePtrReg() + actual_address->constant_value]);
+      }
+      else
+      {
+        m_emit->mov(GetHostReg32(result.host_reg),
+                    m_emit->dword[GetFastmemBasePtrReg() + GetHostReg64(actual_address->host_reg)]);
+      }
+    }
+    break;
+  }
+}
+
 void CodeGenerator::EmitLoadGuestMemoryFastmem(const CodeBlockInstruction& cbi, const Value& address, RegSize size,
                                                Value& result)
 {