SI: Add 16bit accessors for SI IO buffer

Dolphin has traditionally treated the SI IO buffer (128 bytes) as a set of
32 little endian u32s. This works out fine if you only ever read/write
using aligned 32bit accesses. Different sized accesses or misaligned reads
will mess it up. Byte swapping reads/writes will fix this up, but all the
SI devices that use the SI IO buffer need to be fixed [TODO].

Source/Core/Core/HW/SI/SI.cpp

@@ -11,6 +11,7 @@
 
 #include "Common/ChunkFile.h"
 #include "Common/CommonTypes.h"
+#include "Common/Swap.h"
 #include "Core/ConfigManager.h"
 #include "Core/CoreTiming.h"
 #include "Core/HW/MMIO.h"
@@ -50,6 +51,7 @@ enum
   SI_COM_CSR = 0x34,
   SI_STATUS_REG = 0x38,
   SI_EXI_CLOCK_COUNT = 0x3C,
+  SI_IO_BUFFER = 0x80,
 };
 
 // SI Channel Output
@@ -414,12 +416,34 @@ void Shutdown()
 void RegisterMMIO(MMIO::Mapping* mmio, u32 base)
 {
   // Register SI buffer direct accesses.
+  const u32 io_buffer_base = base | SI_IO_BUFFER;
   for (size_t i = 0; i < s_si_buffer.size(); i += sizeof(u32))
   {
-    const u32 address = base | static_cast<u32>(s_si_buffer.size() + i);
+    const u32 address = base | static_cast<u32>(io_buffer_base + i);
 
-    mmio->Register(address, MMIO::DirectRead<u32>((u32*)&s_si_buffer[i]),
-                   MMIO::DirectWrite<u32>((u32*)&s_si_buffer[i]));
+    mmio->Register(address, MMIO::ComplexRead<u32>([i](u32) {
+                     u32 val;
+                     std::memcpy(&val, &s_si_buffer[i], sizeof(val));
+                     return Common::swap32(val);
+                   }),
+                   MMIO::ComplexWrite<u32>([i](u32, u32 val) {
+                     val = Common::swap32(val);
+                     std::memcpy(&s_si_buffer[i], &val, sizeof(val));
+                   }));
+  }
+  for (size_t i = 0; i < s_si_buffer.size(); i += sizeof(u16))
+  {
+    const u32 address = base | static_cast<u32>(io_buffer_base + i);
+
+    mmio->Register(address, MMIO::ComplexRead<u16>([i](u32) {
+                     u16 val;
+                     std::memcpy(&val, &s_si_buffer[i], sizeof(val));
+                     return Common::swap16(val);
+                   }),
+                   MMIO::ComplexWrite<u16>([i](u32, u16 val) {
+                     val = Common::swap16(val);
+                     std::memcpy(&s_si_buffer[i], &val, sizeof(val));
+                   }));
   }
 
   // In and out for the 4 SI channels.

Source/Core/Core/HW/SI/SI_DeviceGCController.cpp

@@ -10,6 +10,7 @@
 #include "Common/CommonTypes.h"
 #include "Common/Logging/Log.h"
 #include "Common/MsgHandler.h"
+#include "Common/Swap.h"
 #include "Core/CoreTiming.h"
 #include "Core/HW/GCPad.h"
 #include "Core/HW/ProcessorInterface.h"
@@ -50,21 +51,21 @@ int CSIDevice_GCController::RunBuffer(u8* buffer, int length)
   GCPadStatus pad_status = GetPadStatus();
   if (!pad_status.isConnected)
   {
-    constexpr u32 reply = SI_ERROR_NO_RESPONSE;
+    u32 reply = Common::swap32(SI_ERROR_NO_RESPONSE);
     std::memcpy(buffer, &reply, sizeof(reply));
     return 4;
   }
 
   // Read the command
-  EBufferCommands command = static_cast<EBufferCommands>(buffer[3]);
+  EBufferCommands command = static_cast<EBufferCommands>(buffer[0]);
 
   // Handle it
   switch (command)
   {
   case CMD_RESET:
   case CMD_ID:
   {
-    constexpr u32 id = SI_GC_CONTROLLER;
+    u32 id = Common::swap32(SI_GC_CONTROLLER);
     std::memcpy(buffer, &id, sizeof(id));
     break;
   }
@@ -76,8 +77,8 @@ int CSIDevice_GCController::RunBuffer(u8* buffer, int length)
     GetData(high, low);
     for (int i = 0; i < (length - 1) / 2; i++)
     {
-      buffer[i + 0] = (high >> (i * 8)) & 0xff;
-      buffer[i + 4] = (low >> (i * 8)) & 0xff;
+      buffer[i + 0] = (high >> (24 - (i * 8))) & 0xff;
+      buffer[i + 4] = (low >> (24 - (i * 8))) & 0xff;
     }
   }
   break;
@@ -92,7 +93,7 @@ int CSIDevice_GCController::RunBuffer(u8* buffer, int length)
     u8* calibration = reinterpret_cast<u8*>(&m_origin);
     for (int i = 0; i < (int)sizeof(SOrigin); i++)
     {
-      buffer[i ^ 3] = *calibration++;
+      buffer[i] = *calibration++;
     }
   }
   break;
@@ -108,7 +109,7 @@ int CSIDevice_GCController::RunBuffer(u8* buffer, int length)
     u8* calibration = reinterpret_cast<u8*>(&m_origin);
     for (int i = 0; i < (int)sizeof(SOrigin); i++)
     {
-      buffer[i ^ 3] = *calibration++;
+      buffer[i] = *calibration++;
     }
   }
   break;