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 adjusted.

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_Device.cpp

@@ -47,7 +47,7 @@ int ISIDevice::RunBuffer(u8* buffer, int length)
 
   while (num < length)
   {
-    temp += StringFromFormat("0x%02x ", buffer[num ^ 3]);
+    temp += StringFromFormat("0x%02x ", buffer[num]);
     num++;
 
     if ((num % 8) == 0)

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

@@ -7,6 +7,7 @@
 #include <cstring>
 
 #include "Common/CommonTypes.h"
+#include "Common/Swap.h"
 #include "InputCommon/GCPadStatus.h"
 
 namespace SerialInterface
@@ -19,13 +20,13 @@ CSIDevice_DanceMat::CSIDevice_DanceMat(SIDevices device, int device_number)
 int CSIDevice_DanceMat::RunBuffer(u8* buffer, int length)
 {
   // Read the command
-  EBufferCommands command = static_cast<EBufferCommands>(buffer[3]);
+  EBufferCommands command = static_cast<EBufferCommands>(buffer[0]);
 
   if (command == CMD_RESET)
   {
     ISIDevice::RunBuffer(buffer, length);
 
-    constexpr u32 id = SI_DANCEMAT;
+    u32 id = Common::swap32(SI_DANCEMAT);
     std::memcpy(buffer, &id, sizeof(id));
   }
   else

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

@@ -16,6 +16,7 @@
 #include "Common/CommonTypes.h"
 #include "Common/Flag.h"
 #include "Common/Logging/Log.h"
+#include "Common/Swap.h"
 #include "Common/Thread.h"
 #include "Core/CoreTiming.h"
 #include "Core/HW/SI/SI_Device.h"
@@ -32,7 +33,7 @@ std::mutex s_cs_gba;
 std::mutex s_cs_gba_clk;
 int s_num_connected;
 Common::Flag s_server_running;
-}
+}  // namespace
 
 enum EJoybusCmds
 {
@@ -235,7 +236,7 @@ void GBASockServer::Send(const u8* si_buffer)
 
   std::array<u8, SEND_MAX_SIZE> send_data;
   for (size_t i = 0; i < send_data.size(); i++)
-    send_data[i] = si_buffer[i ^ 3];
+    send_data[i] = si_buffer[i];
 
   u8 cmd = send_data[0];
   if (cmd != CMD_STATUS)
@@ -281,7 +282,7 @@ int GBASockServer::Receive(u8* si_buffer)
   }
 
   for (size_t i = 0; i < recv_data.size(); i++)
-    si_buffer[i ^ 3] = recv_data[i];
+    si_buffer[i] = recv_data[i];
   return static_cast<int>(std::min(num_received, recv_data.size()));
 }
 
@@ -299,23 +300,23 @@ int CSIDevice_GBA::RunBuffer(u8* buffer, int length)
     if (m_sock_server.Connect())
     {
 #ifdef _DEBUG
-      NOTICE_LOG(SERIALINTERFACE, "%01d cmd %02x [> %02x%02x%02x%02x]", m_device_number, buffer[3],
-                 buffer[2], buffer[1], buffer[0], buffer[7]);
+      NOTICE_LOG(SERIALINTERFACE, "%01d cmd %02x [> %02x%02x%02x%02x]", m_device_number, buffer[0],
+                 buffer[1], buffer[2], buffer[3], buffer[4]);
 #endif
       m_sock_server.Send(buffer);
     }
     else
     {
-      constexpr u32 reply = SI_ERROR_NO_RESPONSE;
+      u32 reply = Common::swap32(SI_ERROR_NO_RESPONSE);
       std::memcpy(buffer, &reply, sizeof(reply));
       return sizeof(reply);
     }
-    m_last_cmd = buffer[3];
+    m_last_cmd = buffer[0];
     m_timestamp_sent = CoreTiming::GetTicks();
     m_next_action = NextAction::WaitTransferTime;
   }
 
-  // [[fallthrough]]
+  // [[fallthrough]]b
   case NextAction::WaitTransferTime:
   {
     int elapsed_time = static_cast<int>(CoreTiming::GetTicks() - m_timestamp_sent);
@@ -332,7 +333,7 @@ int CSIDevice_GBA::RunBuffer(u8* buffer, int length)
     m_next_action = NextAction::SendCommand;
     if (num_data_received == 0)
     {
-      constexpr u32 reply = SI_ERROR_NO_RESPONSE;
+      u32 reply = Common::swap32(SI_ERROR_NO_RESPONSE);
       std::memcpy(buffer, &reply, sizeof(reply));
       return sizeof(reply);
     }
@@ -342,7 +343,7 @@ int CSIDevice_GBA::RunBuffer(u8* buffer, int length)
                                          LogTypes::LWARNING;
     GENERIC_LOG(LogTypes::SERIALINTERFACE, log_level,
                 "%01d                              [< %02x%02x%02x%02x%02x] (%i)", m_device_number,
-                buffer[3], buffer[2], buffer[1], buffer[0], buffer[7], num_data_received);
+                buffer[0], buffer[1], buffer[2], buffer[3], buffer[4], num_data_received);
 #endif
     return num_data_received;
   }

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;

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

@@ -8,6 +8,7 @@
 
 #include "Common/CommonTypes.h"
 #include "Common/Logging/Log.h"
+#include "Common/Swap.h"
 #include "Core/HW/GCPad.h"
 
 namespace SerialInterface
@@ -23,15 +24,15 @@ int CSIDevice_GCSteeringWheel::RunBuffer(u8* buffer, int length)
   ISIDevice::RunBuffer(buffer, length);
 
   // 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_STEERING;
+    u32 id = Common::swap32(SI_GC_STEERING);
     std::memcpy(buffer, &id, sizeof(id));
     break;
   }

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

@@ -9,6 +9,7 @@
 #include "Common/ChunkFile.h"
 #include "Common/CommonTypes.h"
 #include "Common/Logging/Log.h"
+#include "Common/Swap.h"
 #include "Core/HW/GCKeyboard.h"
 #include "InputCommon/KeyboardStatus.h"
 
@@ -26,15 +27,15 @@ int CSIDevice_Keyboard::RunBuffer(u8* buffer, int length)
   ISIDevice::RunBuffer(buffer, length);
 
   // 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_KEYBOARD;
+    u32 id = Common::swap32(SI_GC_KEYBOARD);
     std::memcpy(buffer, &id, sizeof(id));
     break;
   }
@@ -46,8 +47,8 @@ int CSIDevice_Keyboard::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;

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

@@ -3,6 +3,7 @@
 // Refer to the license.txt file included.
 
 #include "Core/HW/SI/SI_DeviceNull.h"
+#include "Common/Swap.h"
 
 #include <cstring>
 
@@ -15,7 +16,7 @@ CSIDevice_Null::CSIDevice_Null(SIDevices device, int device_number)
 
 int CSIDevice_Null::RunBuffer(u8* buffer, int length)
 {
-  constexpr u32 reply = SI_ERROR_NO_RESPONSE;
+  u32 reply = Common::swap32(SI_ERROR_NO_RESPONSE);
   std::memcpy(buffer, &reply, sizeof(reply));
   return 4;
 }