NWRAM & Direct Boot (#1149)

* SCFG_BIOS now selects between NDS and DSi BIOS
Allowing experimental direct boot.

* - making NWRAM Priorities work as in the HW
- fixed loading DSi stage2 Bootloader when in direct boot (should not be loaded) and might interfere with the image loaded into the (N)WRAM previously

* NWRAM and SCFG Registers are now write-guarded by the corresponding SCFG_EXT7/9 Bits

* removed debugging remainder

* Moved NDSHeader dependent SCFG initialization into the new DSi::SetupDirectBoot function called from NDS::SetupDirectBoot when ConsoleType is DSi

* removed redundant SCFG_BIOS bit checks

* Set of changes from RSDuck's review

* removed a forgotten comment

* - removed the guarded debug outputs for NWRAM
- NWRAM writes to bank and window registers now apply their write masks.
- added comment on an existing TODO within the code describing, why this is OK
- fixed initial NWRAM bank reset just accessing set A

* NWRAM not mapped to 03... range if bit25 in SCFG_EXT is cleared

* removed NWRAM write block on cleared bit25 in SCFG_EXT

* changed type of iterator for MapNWRAM_x functions from s8 to int

* - reduced calculations/comparisons on NWRAM write operations
- changed u8 to unsigned int for an iterator MapNWRAM_x

Co-authored-by: Tim Seidel <t.seidel@kts-systeme.de>

3 files changed, 698 insertions, 318 deletions

diff --git a/src/DSi.cpp b/src/DSi.cpp
index b55992c..b22dc8f 100644
--- a/src/DSi.cpp
+++ b/src/DSi.cpp
@@ -162,6 +162,7 @@ void Reset()
     SCFG_EXT[1] = 0x93FFFB06;
     SCFG_MC = 0x0010;//0x0011;
     SCFG_RST = 0;
+
     DSi_DSP::SetRstLine(false);
 
     // LCD init flag
@@ -185,6 +186,39 @@ void Reset()
     ARM7Write16(eaddr+0x42, 0x0001);
 }
 
+void SetupDirectBoot()
+{
+    // If loading a NDS directly, this value should be set
+    // depending on the console type in the header at offset 12h
+    switch (NDSCart::Header.UnitCode)
+    {
+    case 0x00: /* NDS Image */
+        // on a pure NDS Image, we disable all extended features
+        // TODO: For now keep the features enabled, as you can run pure NDS in NDS Emulation anyway
+        SCFG_BIOS = 0x0303;
+        break;
+    case 0x02: /* DSi Enhanced Image */
+        SCFG_BIOS = 0x0303;
+        break;
+    default:
+        SCFG_BIOS = 0x0101; // TODO: should be zero when booting from BIOS
+        break;
+    }
+    // no NWRAM Mapping
+    for (int i = 0; i < 4; i++)
+        MapNWRAM_A(i, 0);
+    for (int i = 0; i < 8; i++)
+        MapNWRAM_B(i, 0);
+    for (int i = 0; i < 8; i++)
+        MapNWRAM_C(i, 0);
+    // No NWRAM Window
+    for (int i = 0; i < 3; i++)
+    {
+        MapNWRAMRange(0, i, 0);
+        MapNWRAMRange(1, i, 0);
+    }
+}
+
 void SoftReset()
 {
     // TODO: check exactly what is reset
@@ -572,6 +606,10 @@ void StopNDMAs(u32 cpu, u32 mode)
 
 void MapNWRAM_A(u32 num, u8 val)
 {
+    // NWRAM Bank A does not allow all bits to be set
+    // possible non working combinations are caught by later code, but these are not set-able at all
+    val &= ~0x72;
+
     if (MBK[0][8] & (1 << num))
     {
         printf("trying to map NWRAM_A %d to %02X, but it is write-protected (%08X)\n", num, val, MBK[0][8]);
@@ -591,22 +629,35 @@ void MapNWRAM_A(u32 num, u8 val)
     MBK[0][mbkn] |= (val << mbks);
     MBK[1][mbkn] = MBK[0][mbkn];
 
-    u8* ptr = &NWRAM_A[num << 16];
-
-    if (oldval & 0x80)
+    // When we only update the mapping on the written MBK, we will
+    // have priority of the last witten MBK over the others
+    // However the hardware has a fixed order. Therefor 
+    // we need to iterate through them all in a fixed order and update
+    // the mapping, so the result is independend on the MBK write order
+    for (unsigned int part = 0; part < 4; part++)
     {
-        if (NWRAMMap_A[oldval & 0x01][(oldval >> 2) & 0x3] == ptr)
-            NWRAMMap_A[oldval & 0x01][(oldval >> 2) & 0x3] = NULL;
+        NWRAMMap_A[0][part] = NULL;
+        NWRAMMap_A[1][part] = NULL;
     }
-
-    if (val & 0x80)
+    for (int part = 3; part >= 0; part--)
     {
-        NWRAMMap_A[val & 0x01][(val >> 2) & 0x3] = ptr;
+        u8* ptr = &NWRAM_A[part << 16];
+
+        if ((MBK[0][0 + (part / 4)] >> ((part % 4) * 8)) & 0x80)
+        {
+            u8 mVal = (MBK[0][0 + (part / 4)] >> ((part % 4) * 8)) & 0xfd;
+
+            NWRAMMap_A[mVal & 0x03][(mVal >> 2) & 0x3] = ptr;
+        }
     }
 }
 
 void MapNWRAM_B(u32 num, u8 val)
 {
+    // NWRAM Bank B does not allow all bits to be set
+    // possible non working combinations are caught by later code, but these are not set-able at all
+    val &= ~0x60;
+
     if (MBK[0][8] & (1 << (8+num)))
     {
         printf("trying to map NWRAM_B %d to %02X, but it is write-protected (%08X)\n", num, val, MBK[0][8]);
@@ -626,28 +677,42 @@ void MapNWRAM_B(u32 num, u8 val)
     MBK[0][mbkn] |= (val << mbks);
     MBK[1][mbkn] = MBK[0][mbkn];
 
-    u8* ptr = &NWRAM_B[num << 15];
-
-    DSi_DSP::OnMBKCfg('B', num, oldval, val, ptr);
-
-    if (oldval & 0x80)
+    // When we only update the mapping on the written MBK, we will
+    // have priority of the last witten MBK over the others
+    // However the hardware has a fixed order. Therefor 
+    // we need to iterate through them all in a fixed order and update
+    // the mapping, so the result is independend on the MBK write order
+    for (unsigned int part = 0; part < 8; part++)
     {
-        if (oldval & 0x02) oldval &= 0xFE;
-
-        if (NWRAMMap_B[oldval & 0x03][(oldval >> 2) & 0x7] == ptr)
-            NWRAMMap_B[oldval & 0x03][(oldval >> 2) & 0x7] = NULL;
+        NWRAMMap_B[0][part] = NULL;
+        NWRAMMap_B[1][part] = NULL;
+        NWRAMMap_B[2][part] = NULL;
     }
-
-    if (val & 0x80)
+    for (int part = 7; part >= 0; part--)
     {
-        if (val & 0x02) val &= 0xFE;
+        u8* ptr = &NWRAM_B[part << 15];
+
+        if (part == num)
+        {
+            DSi_DSP::OnMBKCfg('B', num, oldval, val, ptr);
+        }
 
-        NWRAMMap_B[val & 0x03][(val >> 2) & 0x7] = ptr;
+        if ((MBK[0][1 + (part / 4)] >> ((part % 4) * 8)) & 0x80)
+        {
+            u8 mVal = (MBK[0][1 + (part / 4)] >> ((part % 4) * 8)) & 0xff;
+            if (mVal & 0x02) mVal &= 0xFE;
+
+            NWRAMMap_B[mVal & 0x03][(mVal >> 2) & 0x7] = ptr;
+        }
     }
 }
 
 void MapNWRAM_C(u32 num, u8 val)
 {
+    // NWRAM Bank C does not allow all bits to be set
+    // possible non working combinations are caught by later code, but these are not set-able at all
+    val &= ~0x60;
+
     if (MBK[0][8] & (1 << (16+num)))
     {
         printf("trying to map NWRAM_C %d to %02X, but it is write-protected (%08X)\n", num, val, MBK[0][8]);
@@ -667,28 +732,51 @@ void MapNWRAM_C(u32 num, u8 val)
     MBK[0][mbkn] |= (val << mbks);
     MBK[1][mbkn] = MBK[0][mbkn];
 
-    u8* ptr = &NWRAM_C[num << 15];
-
-    DSi_DSP::OnMBKCfg('C', num, oldval, val, ptr);
-
-    if (oldval & 0x80)
+    // When we only update the mapping on the written MBK, we will
+    // have priority of the last witten MBK over the others
+    // However the hardware has a fixed order. Therefor 
+    // we need to iterate through them all in a fixed order and update
+    // the mapping, so the result is independend on the MBK write order
+    for (unsigned int part = 0; part < 8; part++)
     {
-        if (oldval & 0x02) oldval &= 0xFE;
-
-        if (NWRAMMap_C[oldval & 0x03][(oldval >> 2) & 0x7] == ptr)
-            NWRAMMap_C[oldval & 0x03][(oldval >> 2) & 0x7] = NULL;
+        NWRAMMap_C[0][part] = NULL;
+        NWRAMMap_C[1][part] = NULL;
+        NWRAMMap_C[2][part] = NULL;
     }
-
-    if (val & 0x80)
+    for (int part = 7; part >= 0; part--)
     {
-        if (val & 0x02) val &= 0xFE;
+        u8* ptr = &NWRAM_C[part << 15];
+
+        if (part == num)
+        {
+            DSi_DSP::OnMBKCfg('C', num, oldval, val, ptr);
+        }
 
-        NWRAMMap_C[val & 0x03][(val >> 2) & 0x7] = ptr;
+        if ((MBK[0][3 + (part / 4)] >> ((part % 4) * 8)) & 0x80)
+        {
+            u8 mVal = MBK[0][3 + (part / 4)] >> ((part % 4) *8) & 0xff;
+            if (mVal & 0x02) mVal &= 0xFE;
+            NWRAMMap_C[mVal & 0x03][(mVal >> 2) & 0x7] = ptr;
+        }
     }
 }
 
 void MapNWRAMRange(u32 cpu, u32 num, u32 val)
 {
+    // The windowing registers are not writeable in all bits
+    // We need to do this before the change test, so we do not
+    // get false "was changed" fall throughs
+    switch (num)
+    {
+        case 0:
+            val &= ~0xE00FC00F;
+            break;
+        case 1:
+        case 2:
+            val &= ~0xE007C007;
+            break;
+    }
+
     u32 oldval = MBK[cpu][5+num];
     if (oldval == val) return;
 
@@ -698,7 +786,14 @@ void MapNWRAMRange(u32 cpu, u32 num, u32 val)
 
     MBK[cpu][5+num] = val;
 
-    // TODO: what happens when the ranges are 'out of range'????
+    // Was TODO: What happens when the ranges are 'out of range'????
+    // Answer: The actual range is limited to 0x03000000 to 0x03ffffff
+    // The NWRAM can not map into the HW Register ranges and is never
+    // mapped there. However the end indizes are allowed to have a value
+    // exceeding that range, in which the accessed area is just cut
+    // at the end of the region
+    // since melonDS does this cut by the case switch in the read/write
+    // functions, we do not need to care here.
     if (num == 0)
     {
         u32 start = 0x03000000 + (((val >> 4) & 0xFF) << 16);
@@ -802,20 +897,24 @@ u8 ARM9Read8(u32 addr)
     switch (addr & 0xFF000000)
     {
     case 0x03000000:
-        if (addr >= NWRAMStart[0][0] && addr < NWRAMEnd[0][0])
+    case 0x03800000:
+        if (SCFG_EXT[0] & (1 << 25))
         {
-            u8* ptr = NWRAMMap_A[0][(addr >> 16) & NWRAMMask[0][0]];
-            return ptr ? *(u8*)&ptr[addr & 0xFFFF] : 0;
-        }
-        if (addr >= NWRAMStart[0][1] && addr < NWRAMEnd[0][1])
-        {
-            u8* ptr = NWRAMMap_B[0][(addr >> 15) & NWRAMMask[0][1]];
-            return ptr ? *(u8*)&ptr[addr & 0x7FFF] : 0;
-        }
-        if (addr >= NWRAMStart[0][2] && addr < NWRAMEnd[0][2])
-        {
-            u8* ptr = NWRAMMap_C[0][(addr >> 15) & NWRAMMask[0][2]];
-            return ptr ? *(u8*)&ptr[addr & 0x7FFF] : 0;
+            if (addr >= NWRAMStart[0][0] && addr < NWRAMEnd[0][0])
+            {
+                u8* ptr = NWRAMMap_A[0][(addr >> 16) & NWRAMMask[0][0]];
+                return ptr ? *(u8*)&ptr[addr & 0xFFFF] : 0;
+            }
+            if (addr >= NWRAMStart[0][1] && addr < NWRAMEnd[0][1])
+            {
+                u8* ptr = NWRAMMap_B[0][(addr >> 15) & NWRAMMask[0][1]];
+                return ptr ? *(u8*)&ptr[addr & 0x7FFF] : 0;
+            }
+            if (addr >= NWRAMStart[0][2] && addr < NWRAMEnd[0][2])
+            {
+                u8* ptr = NWRAMMap_C[0][(addr >> 15) & NWRAMMask[0][2]];
+                return ptr ? *(u8*)&ptr[addr & 0x7FFF] : 0;
+            }
         }
         return NDS::ARM9Read8(addr);
 
@@ -844,20 +943,24 @@ u16 ARM9Read16(u32 addr)
     switch (addr & 0xFF000000)
     {
     case 0x03000000:
-        if (addr >= NWRAMStart[0][0] && addr < NWRAMEnd[0][0])
-        {
-            u8* ptr = NWRAMMap_A[0][(addr >> 16) & NWRAMMask[0][0]];
-            return ptr ? *(u16*)&ptr[addr & 0xFFFF] : 0;
-        }
-        if (addr >= NWRAMStart[0][1] && addr < NWRAMEnd[0][1])
-        {
-            u8* ptr = NWRAMMap_B[0][(addr >> 15) & NWRAMMask[0][1]];
-            return ptr ? *(u16*)&ptr[addr & 0x7FFF] : 0;
-        }
-        if (addr >= NWRAMStart[0][2] && addr < NWRAMEnd[0][2])
+    case 0x03800000:
+        if (SCFG_EXT[0] & (1 << 25))
         {
-            u8* ptr = NWRAMMap_C[0][(addr >> 15) & NWRAMMask[0][2]];
-            return ptr ? *(u16*)&ptr[addr & 0x7FFF] : 0;
+            if (addr >= NWRAMStart[0][0] && addr < NWRAMEnd[0][0])
+            {
+                u8* ptr = NWRAMMap_A[0][(addr >> 16) & NWRAMMask[0][0]];
+                return ptr ? *(u16*)&ptr[addr & 0xFFFF] : 0;
+            }
+            if (addr >= NWRAMStart[0][1] && addr < NWRAMEnd[0][1])
+            {
+                u8* ptr = NWRAMMap_B[0][(addr >> 15) & NWRAMMask[0][1]];
+                return ptr ? *(u16*)&ptr[addr & 0x7FFF] : 0;
+            }
+            if (addr >= NWRAMStart[0][2] && addr < NWRAMEnd[0][2])
+            {
+                u8* ptr = NWRAMMap_C[0][(addr >> 15) & NWRAMMask[0][2]];
+                return ptr ? *(u16*)&ptr[addr & 0x7FFF] : 0;
+            }
         }
         return NDS::ARM9Read16(addr);
 
@@ -879,7 +982,6 @@ u32 ARM9Read32(u32 addr)
     {
         if ((addr >= 0xFFFF8000) && (SCFG_BIOS & (1<<0)))
             return 0xFFFFFFFF;
-
         return *(u32*)&ARM9iBIOS[addr & 0xFFFF];
     }
 
@@ -892,20 +994,24 @@ u32 ARM9Read32(u32 addr)
         break;
 
     case 0x03000000:
-        if (addr >= NWRAMStart[0][0] && addr < NWRAMEnd[0][0])
+    case 0x03800000:
+        if (SCFG_EXT[0] & (1 << 25))
         {
-            u8* ptr = NWRAMMap_A[0][(addr >> 16) & NWRAMMask[0][0]];
-            return ptr ? *(u32*)&ptr[addr & 0xFFFF] : 0;
-        }
-        if (addr >= NWRAMStart[0][1] && addr < NWRAMEnd[0][1])
-        {
-            u8* ptr = NWRAMMap_B[0][(addr >> 15) & NWRAMMask[0][1]];
-            return ptr ? *(u32*)&ptr[addr & 0x7FFF] : 0;
-        }
-        if (addr >= NWRAMStart[0][2] && addr < NWRAMEnd[0][2])
-        {
-            u8* ptr = NWRAMMap_C[0][(addr >> 15) & NWRAMMask[0][2]];
-            return ptr ? *(u32*)&ptr[addr & 0x7FFF] : 0;
+            if (addr >= NWRAMStart[0][0] && addr < NWRAMEnd[0][0])
+            {
+                u8* ptr = NWRAMMap_A[0][(addr >> 16) & NWRAMMask[0][0]];
+                return ptr ? *(u32*)&ptr[addr & 0xFFFF] : 0;
+            }
+            if (addr >= NWRAMStart[0][1] && addr < NWRAMEnd[0][1])
+            {
+                u8* ptr = NWRAMMap_B[0][(addr >> 15) & NWRAMMask[0][1]];
+                return ptr ? *(u32*)&ptr[addr & 0x7FFF] : 0;
+            }
+            if (addr >= NWRAMStart[0][2] && addr < NWRAMEnd[0][2])
+            {
+                u8* ptr = NWRAMMap_C[0][(addr >> 15) & NWRAMMask[0][2]];
+                return ptr ? *(u32*)&ptr[addr & 0x7FFF] : 0;
+            }
         }
         return NDS::ARM9Read32(addr);
 
@@ -926,41 +1032,69 @@ void ARM9Write8(u32 addr, u8 val)
     switch (addr & 0xFF000000)
     {
     case 0x03000000:
-        if (addr >= NWRAMStart[0][0] && addr < NWRAMEnd[0][0])
+    case 0x03800000:
+        if (SCFG_EXT[0] & (1 << 25))
         {
-            u8* ptr = NWRAMMap_A[0][(addr >> 16) & NWRAMMask[0][0]];
-            if (ptr)
+            if (addr >= NWRAMStart[0][0] && addr < NWRAMEnd[0][0])
             {
-                *(u8*)&ptr[addr & 0xFFFF] = val;
+                // Write to a bank is special, as it writes to all
+                // parts that are mapped and not just the highest priority
+                // See http://melonds.kuribo64.net/board/thread.php?pid=3974#3974
+                // so we need to iterate through all parts and write to all mapped here
+                unsigned int destPartSetting = ((addr >> 13) & (NWRAMMask[0][0] << 3)) | 0x80;
+                for (int page = 0; page < 4; page++)
+                {
+                    unsigned int bankInfo = (MBK[0][0 + (page / 4)] >> ((page % 4) * 8)) & 0xff;
+                    if (bankInfo != destPartSetting)
+                        continue;
+                    u8* ptr = &NWRAM_A[page * 0x8000];
+                    *(u8*)&ptr[addr & 0x7FFF] = val;
 #ifdef JIT_ENABLED
-                ARMJIT::CheckAndInvalidate<0, ARMJIT_Memory::memregion_NewSharedWRAM_A>(addr);
+                    ARMJIT::CheckAndInvalidate<0, ARMJIT_Memory::memregion_NewSharedWRAM_A>(addr);
 #endif
+                }
+                return;
             }
-            return;
-        }
-        if (addr >= NWRAMStart[0][1] && addr < NWRAMEnd[0][1])
-        {
-            u8* ptr = NWRAMMap_B[0][(addr >> 15) & NWRAMMask[0][1]];
-            if (ptr)
+            if (addr >= NWRAMStart[0][1] && addr < NWRAMEnd[0][1])
             {
-                *(u8*)&ptr[addr & 0x7FFF] = val;
+                // Write to a bank is special, as it writes to all
+                // parts that are mapped and not just the highest priority
+                // See http://melonds.kuribo64.net/board/thread.php?pid=3974#3974
+                // so we need to iterate through all parts and write to all mapped here
+                unsigned int destPartSetting = ((addr >> 13) & (NWRAMMask[0][1] << 2)) | 0x80;
+                for (int page = 0; page < 8; page++)
+                {
+                    unsigned int bankInfo = (MBK[0][1 + (page / 4)] >> ((page % 4) * 8)) & 0xff;
+                    if (bankInfo != destPartSetting)
+                        continue;
+                    u8* ptr = &NWRAM_B[page * 0x8000];
+                    *(u8*)&ptr[addr & 0x7FFF] = val;
 #ifdef JIT_ENABLED
-                ARMJIT::CheckAndInvalidate<0, ARMJIT_Memory::memregion_NewSharedWRAM_B>(addr);
+                    ARMJIT::CheckAndInvalidate<0, ARMJIT_Memory::memregion_NewSharedWRAM_B>(addr);
 #endif
+                }
+                return;
             }
-            return;
-        }
-        if (addr >= NWRAMStart[0][2] && addr < NWRAMEnd[0][2])
-        {
-            u8* ptr = NWRAMMap_C[0][(addr >> 15) & NWRAMMask[0][2]];
-            if (ptr)
+            if (addr >= NWRAMStart[0][2] && addr < NWRAMEnd[0][2])
             {
-                *(u8*)&ptr[addr & 0x7FFF] = val;
+                // Write to a bank is special, as it writes to all
+                // parts that are mapped and not just the highest priority
+                // See http://melonds.kuribo64.net/board/thread.php?pid=3974#3974
+                // so we need to iterate through all parts and write to all mapped here
+                unsigned int destPartSetting = ((addr >> 13) & (NWRAMMask[0][2] << 2)) | 0x80;
+                for (int page = 0; page < 8; page++)
+                {
+                    unsigned int bankInfo = (MBK[0][3 + (page / 4)] >> ((page % 4) * 8)) & 0xff;
+                    if (bankInfo != destPartSetting)
+                        continue;
+                    u8* ptr = &NWRAM_C[page * 0x8000];
+                    *(u8*)&ptr[addr & 0x7FFF] = val;
 #ifdef JIT_ENABLED
-                ARMJIT::CheckAndInvalidate<0, ARMJIT_Memory::memregion_NewSharedWRAM_C>(addr);
+                    ARMJIT::CheckAndInvalidate<0, ARMJIT_Memory::memregion_NewSharedWRAM_C>(addr);
 #endif
+                }
+                return;
             }
-            return;
         }
         return NDS::ARM9Write8(addr, val);
 
@@ -996,41 +1130,69 @@ void ARM9Write16(u32 addr, u16 val)
     switch (addr & 0xFF000000)
     {
     case 0x03000000:
-        if (addr >= NWRAMStart[0][0] && addr < NWRAMEnd[0][0])
+    case 0x03800000:
+        if (SCFG_EXT[0] & (1 << 25))
         {
-            u8* ptr = NWRAMMap_A[0][(addr >> 16) & NWRAMMask[0][0]];
-            if (ptr)
+            if (addr >= NWRAMStart[0][0] && addr < NWRAMEnd[0][0])
             {
-                *(u16*)&ptr[addr & 0xFFFF] = val;
+                // Write to a bank is special, as it writes to all
+                // parts that are mapped and not just the highest priority
+                // See http://melonds.kuribo64.net/board/thread.php?pid=3974#3974
+                // so we need to iterate through all parts and write to all mapped here
+                unsigned int destPartSetting = ((addr >> 13) & (NWRAMMask[0][0] << 3)) | 0x80;
+                for (int page = 0; page < 4; page++)
+                {
+                    unsigned int bankInfo = (MBK[0][0 + (page / 4)] >> ((page % 4) * 8)) & 0xff;
+                    if (bankInfo != destPartSetting)
+                        continue;
+                    u8* ptr = &NWRAM_A[page * 0x8000];
+                    *(u16*)&ptr[addr & 0x7FFF] = val;
 #ifdef JIT_ENABLED
-                ARMJIT::CheckAndInvalidate<0, ARMJIT_Memory::memregion_NewSharedWRAM_A>(addr);
+                    ARMJIT::CheckAndInvalidate<0, ARMJIT_Memory::memregion_NewSharedWRAM_A>(addr);
 #endif
+                }
+                return;
             }
-            return;
-        }
-        if (addr >= NWRAMStart[0][1] && addr < NWRAMEnd[0][1])
-        {
-            u8* ptr = NWRAMMap_B[0][(addr >> 15) & NWRAMMask[0][1]];
-            if (ptr)
+            if (addr >= NWRAMStart[0][1] && addr < NWRAMEnd[0][1])
             {
-                *(u16*)&ptr[addr & 0x7FFF] = val;
+                // Write to a bank is special, as it writes to all
+                // parts that are mapped and not just the highest priority
+                // See http://melonds.kuribo64.net/board/thread.php?pid=3974#3974
+                // so we need to iterate through all parts and write to all mapped here
+                unsigned int destPartSetting = ((addr >> 13) & (NWRAMMask[0][1] << 2)) | 0x80;
+                for (int page = 0; page < 8; page++)
+                {
+                    unsigned int bankInfo = (MBK[0][1 + (page / 4)] >> ((page % 4) * 8)) & 0xff;
+                    if (bankInfo != destPartSetting)
+                        continue;
+                    u8* ptr = &NWRAM_B[page * 0x8000];
+                    *(u16*)&ptr[addr & 0x7FFF] = val;
 #ifdef JIT_ENABLED
-                ARMJIT::CheckAndInvalidate<0, ARMJIT_Memory::memregion_NewSharedWRAM_B>(addr);
+                    ARMJIT::CheckAndInvalidate<0, ARMJIT_Memory::memregion_NewSharedWRAM_B>(addr);
 #endif
+                }
+                return;
             }
-            return;
-        }
-        if (addr >= NWRAMStart[0][2] && addr < NWRAMEnd[0][2])
-        {
-            u8* ptr = NWRAMMap_C[0][(addr >> 15) & NWRAMMask[0][2]];
-            if (ptr)
+            if (addr >= NWRAMStart[0][2] && addr < NWRAMEnd[0][2])
             {
-                *(u16*)&ptr[addr & 0x7FFF] = val;
+                // Write to a bank is special, as it writes to all
+                // parts that are mapped and not just the highest priority
+                // See http://melonds.kuribo64.net/board/thread.php?pid=3974#3974
+                // so we need to iterate through all parts and write to all mapped here
+                unsigned int destPartSetting = ((addr >> 13) & (NWRAMMask[0][2] << 2)) | 0x80;
+                for (int page = 0; page < 8; page++)
+                {
+                    unsigned int bankInfo = (MBK[0][3 + (page / 4)] >> ((page % 4) * 8)) & 0xff;
+                    if (bankInfo != destPartSetting)
+                        continue;
+                    u8* ptr = &NWRAM_C[page * 0x8000];
+                    *(u16*)&ptr[addr & 0x7FFF] = val;
 #ifdef JIT_ENABLED
-                ARMJIT::CheckAndInvalidate<0, ARMJIT_Memory::memregion_NewSharedWRAM_C>(addr);
+                    ARMJIT::CheckAndInvalidate<0, ARMJIT_Memory::memregion_NewSharedWRAM_C>(addr);
 #endif
+                }
+                return;
             }
-            return;
         }
         return NDS::ARM9Write16(addr, val);
 
@@ -1052,41 +1214,69 @@ void ARM9Write32(u32 addr, u32 val)
     switch (addr & 0xFF000000)
     {
     case 0x03000000:
-        if (addr >= NWRAMStart[0][0] && addr < NWRAMEnd[0][0])
+    case 0x03800000:
+        if (SCFG_EXT[0] & (1 << 25))
         {
-            u8* ptr = NWRAMMap_A[0][(addr >> 16) & NWRAMMask[0][0]];
-            if (ptr)
+            if (addr >= NWRAMStart[0][0] && addr < NWRAMEnd[0][0])
             {
-                *(u32*)&ptr[addr & 0xFFFF] = val;
+                // Write to a bank is special, as it writes to all
+                // parts that are mapped and not just the highest priority
+                // See http://melonds.kuribo64.net/board/thread.php?pid=3974#3974
+                // so we need to iterate through all parts and write to all mapped here
+                unsigned int destPartSetting = ((addr >> 13) & (NWRAMMask[0][0] << 3)) | 0x80;
+                for (int page = 0; page < 4; page++)
+                {
+                    unsigned int bankInfo = (MBK[0][0 + (page / 4)] >> ((page % 4) * 8)) & 0xff;
+                    if (bankInfo != destPartSetting)
+                        continue;
+                    u8* ptr = &NWRAM_A[page * 0x8000];
+                    *(u32*)&ptr[addr & 0x7FFF] = val;
 #ifdef JIT_ENABLED
-                ARMJIT::CheckAndInvalidate<0, ARMJIT_Memory::memregion_NewSharedWRAM_A>(addr);
+                    ARMJIT::CheckAndInvalidate<0, ARMJIT_Memory::memregion_NewSharedWRAM_A>(addr);
 #endif
+                }
+                return;
             }
-            return;
-        }
-        if (addr >= NWRAMStart[0][1] && addr < NWRAMEnd[0][1])
-        {
-            u8* ptr = NWRAMMap_B[0][(addr >> 15) & NWRAMMask[0][1]];
-            if (ptr)
+            if (addr >= NWRAMStart[0][1] && addr < NWRAMEnd[0][1])
             {
-                *(u32*)&ptr[addr & 0x7FFF] = val;
+                // Write to a bank is special, as it writes to all
+                // parts that are mapped and not just the highest priority
+                // See http://melonds.kuribo64.net/board/thread.php?pid=3974#3974
+                // so we need to iterate through all parts and write to all mapped here
+                unsigned int destPartSetting = ((addr >> 13) & (NWRAMMask[0][1] << 2)) | 0x80;
+                for (int page = 0; page < 8; page++)
+                {
+                    unsigned int bankInfo = (MBK[0][1 + (page / 4)] >> ((page % 4) * 8)) & 0xff;
+                    if (bankInfo != destPartSetting)
+                        continue;
+                    u8* ptr = &NWRAM_B[page * 0x8000];
+                    *(u32*)&ptr[addr & 0x7FFF] = val;
 #ifdef JIT_ENABLED
-                ARMJIT::CheckAndInvalidate<0, ARMJIT_Memory::memregion_NewSharedWRAM_B>(addr);
+                    ARMJIT::CheckAndInvalidate<0, ARMJIT_Memory::memregion_NewSharedWRAM_B>(addr);
 #endif
+                }
+                return;
             }
-            return;
-        }
-        if (addr >= NWRAMStart[0][2] && addr < NWRAMEnd[0][2])
-        {
-            u8* ptr = NWRAMMap_C[0][(addr >> 15) & NWRAMMask[0][2]];
-            if (ptr)
+            if (addr >= NWRAMStart[0][2] && addr < NWRAMEnd[0][2])
             {
-                *(u32*)&ptr[addr & 0x7FFF] = val;
+                // Write to a bank is special, as it writes to all
+                // parts that are mapped and not just the highest priority
+                // See http://melonds.kuribo64.net/board/thread.php?pid=3974#3974
+                // so we need to iterate through all parts and write to all mapped here
+                unsigned int destPartSetting = ((addr >> 13) & (NWRAMMask[0][2] << 2)) | 0x80;
+                for (int page = 0; page < 8; page++)
+                {
+                    unsigned int bankInfo = (MBK[0][3 + (page / 4)] >> ((page % 4) * 8)) & 0xff;
+                    if (bankInfo != destPartSetting)
+                        continue;
+                    u8* ptr = &NWRAM_C[page * 0x8000];
+                    *(u32*)&ptr[addr & 0x7FFF] = val;
 #ifdef JIT_ENABLED
-                ARMJIT::CheckAndInvalidate<0, ARMJIT_Memory::memregion_NewSharedWRAM_C>(addr);
+                    ARMJIT::CheckAndInvalidate<0, ARMJIT_Memory::memregion_NewSharedWRAM_C>(addr);
 #endif
+                }
+                return;
             }
-            return;
         }
         return NDS::ARM9Write32(addr, val);
 
@@ -1151,26 +1341,30 @@ u8 ARM7Read8(u32 addr)
         if (addr < NDS::ARM7BIOSProt && NDS::ARM7->R[15] >= NDS::ARM7BIOSProt)
             return 0xFF;
 
-        return *(u8*)&ARM7iBIOS[addr];
+        return *(u8*)&ARM7iBIOS[addr & 0xFFFF];
     }
 
     switch (addr & 0xFF800000)
     {
     case 0x03000000:
-        if (addr >= NWRAMStart[1][0] && addr < NWRAMEnd[1][0])
-        {
-            u8* ptr = NWRAMMap_A[1][(addr >> 16) & NWRAMMask[1][0]];
-            return ptr ? *(u8*)&ptr[addr & 0xFFFF] : 0;
-        }
-        if (addr >= NWRAMStart[1][1] && addr < NWRAMEnd[1][1])
-        {
-            u8* ptr = NWRAMMap_B[1][(addr >> 15) & NWRAMMask[1][1]];
-            return ptr ? *(u8*)&ptr[addr & 0x7FFF] : 0;
-        }
-        if (addr >= NWRAMStart[1][2] && addr < NWRAMEnd[1][2])
+    case 0x03800000:
+        if (SCFG_EXT[1] & (1 << 25))
         {
-            u8* ptr = NWRAMMap_C[1][(addr >> 15) & NWRAMMask[1][2]];
-            return ptr ? *(u8*)&ptr[addr & 0x7FFF] : 0;
+            if (addr >= NWRAMStart[1][0] && addr < NWRAMEnd[1][0])
+            {
+                u8* ptr = NWRAMMap_A[1][(addr >> 16) & NWRAMMask[1][0]];
+                return ptr ? *(u8*)&ptr[addr & 0xFFFF] : 0;
+            }
+            if (addr >= NWRAMStart[1][1] && addr < NWRAMEnd[1][1])
+            {
+                u8* ptr = NWRAMMap_B[1][(addr >> 15) & NWRAMMask[1][1]];
+                return ptr ? *(u8*)&ptr[addr & 0x7FFF] : 0;
+            }
+            if (addr >= NWRAMStart[1][2] && addr < NWRAMEnd[1][2])
+            {
+                u8* ptr = NWRAMMap_C[1][(addr >> 15) & NWRAMMask[1][2]];
+                return ptr ? *(u8*)&ptr[addr & 0x7FFF] : 0;
+            }
         }
         return NDS::ARM7Read8(addr);
 
@@ -1200,26 +1394,30 @@ u16 ARM7Read16(u32 addr)
         if (addr < NDS::ARM7BIOSProt && NDS::ARM7->R[15] >= NDS::ARM7BIOSProt)
             return 0xFFFF;
 
-        return *(u16*)&ARM7iBIOS[addr];
+        return *(u16*)&ARM7iBIOS[addr & 0xFFFF];
     }
 
     switch (addr & 0xFF800000)
     {
     case 0x03000000:
-        if (addr >= NWRAMStart[1][0] && addr < NWRAMEnd[1][0])
-        {
-            u8* ptr = NWRAMMap_A[1][(addr >> 16) & NWRAMMask[1][0]];
-            return ptr ? *(u16*)&ptr[addr & 0xFFFF] : 0;
-        }
-        if (addr >= NWRAMStart[1][1] && addr < NWRAMEnd[1][1])
-        {
-            u8* ptr = NWRAMMap_B[1][(addr >> 15) & NWRAMMask[1][1]];
-            return ptr ? *(u16*)&ptr[addr & 0x7FFF] : 0;
-        }
-        if (addr >= NWRAMStart[1][2] && addr < NWRAMEnd[1][2])
+    case 0x03800000:
+        if (SCFG_EXT[1] & (1 << 25))
         {
-            u8* ptr = NWRAMMap_C[1][(addr >> 15) & NWRAMMask[1][2]];
-            return ptr ? *(u16*)&ptr[addr & 0x7FFF] : 0;
+            if (addr >= NWRAMStart[1][0] && addr < NWRAMEnd[1][0])
+            {
+                u8* ptr = NWRAMMap_A[1][(addr >> 16) & NWRAMMask[1][0]];
+                return ptr ? *(u16*)&ptr[addr & 0xFFFF] : 0;
+            }
+            if (addr >= NWRAMStart[1][1] && addr < NWRAMEnd[1][1])
+            {
+                u8* ptr = NWRAMMap_B[1][(addr >> 15) & NWRAMMask[1][1]];
+                return ptr ? *(u16*)&ptr[addr & 0x7FFF] : 0;
+            }
+            if (addr >= NWRAMStart[1][2] && addr < NWRAMEnd[1][2])
+            {
+                u8* ptr = NWRAMMap_C[1][(addr >> 15) & NWRAMMask[1][2]];
+                return ptr ? *(u16*)&ptr[addr & 0x7FFF] : 0;
+            }
         }
         return NDS::ARM7Read16(addr);
 
@@ -1249,26 +1447,30 @@ u32 ARM7Read32(u32 addr)
         if (addr < NDS::ARM7BIOSProt && NDS::ARM7->R[15] >= NDS::ARM7BIOSProt)
             return 0xFFFFFFFF;
 
-        return *(u32*)&ARM7iBIOS[addr];
+        return *(u32*)&ARM7iBIOS[addr & 0xFFFF];
     }
 
     switch (addr & 0xFF800000)
     {
     case 0x03000000:
-        if (addr >= NWRAMStart[1][0] && addr < NWRAMEnd[1][0])
-        {
-            u8* ptr = NWRAMMap_A[1][(addr >> 16) & NWRAMMask[1][0]];
-            return ptr ? *(u32*)&ptr[addr & 0xFFFF] : 0;
-        }
-        if (addr >= NWRAMStart[1][1] && addr < NWRAMEnd[1][1])
+    case 0x03800000:
+        if (SCFG_EXT[1] & (1 << 25))
         {
-            u8* ptr = NWRAMMap_B[1][(addr >> 15) & NWRAMMask[1][1]];
-            return ptr ? *(u32*)&ptr[addr & 0x7FFF] : 0;
-        }
-        if (addr >= NWRAMStart[1][2] && addr < NWRAMEnd[1][2])
-        {
-            u8* ptr = NWRAMMap_C[1][(addr >> 15) & NWRAMMask[1][2]];
-            return ptr ? *(u32*)&ptr[addr & 0x7FFF] : 0;
+            if (addr >= NWRAMStart[1][0] && addr < NWRAMEnd[1][0])
+            {
+                u8* ptr = NWRAMMap_A[1][(addr >> 16) & NWRAMMask[1][0]];
+                return ptr ? *(u32*)&ptr[addr & 0xFFFF] : 0;
+            }
+            if (addr >= NWRAMStart[1][1] && addr < NWRAMEnd[1][1])
+            {
+                u8* ptr = NWRAMMap_B[1][(addr >> 15) & NWRAMMask[1][1]];
+                return ptr ? *(u32*)&ptr[addr & 0x7FFF] : 0;
+            }
+            if (addr >= NWRAMStart[1][2] && addr < NWRAMEnd[1][2])
+            {
+                u8* ptr = NWRAMMap_C[1][(addr >> 15) & NWRAMMask[1][2]];
+                return ptr ? *(u32*)&ptr[addr & 0x7FFF] : 0;
+            }
         }
         return NDS::ARM7Read32(addr);
 
@@ -1292,41 +1494,69 @@ void ARM7Write8(u32 addr, u8 val)
     switch (addr & 0xFF800000)
     {
     case 0x03000000:
-        if (addr >= NWRAMStart[1][0] && addr < NWRAMEnd[1][0])
+    case 0x03800000:
+        if (SCFG_EXT[1] & (1 << 25))
         {
-            u8* ptr = NWRAMMap_A[1][(addr >> 16) & NWRAMMask[1][0]];
-            if (ptr)
+            if (addr >= NWRAMStart[1][0] && addr < NWRAMEnd[1][0])
             {
-                *(u8*)&ptr[addr & 0xFFFF] = val;
+                // Write to a bank is special, as it writes to all
+                // parts that are mapped and not just the highest priority
+                // See http://melonds.kuribo64.net/board/thread.php?pid=3974#3974
+                // so we need to iterate through all parts and write to all mapped here
+                unsigned int destPartSetting = ((addr >> 13) & (NWRAMMask[1][0] << 3)) | 0x81;
+                for (int page = 0; page < 4; page++)
+                {
+                    unsigned int bankInfo = (MBK[1][0 + (page / 4)] >> ((page % 4) * 8)) & 0xff;
+                    if (bankInfo != destPartSetting)
+                        continue;
+                    u8* ptr = &NWRAM_A[page * 0x8000];
+                    *(u8*)&ptr[addr & 0x7FFF] = val;
 #ifdef JIT_ENABLED
-                ARMJIT::CheckAndInvalidate<1, ARMJIT_Memory::memregion_NewSharedWRAM_A>(addr);
+                    ARMJIT::CheckAndInvalidate<1, ARMJIT_Memory::memregion_NewSharedWRAM_A>(addr);
 #endif
+                }
+                return;
             }
-            return;
-        }
-        if (addr >= NWRAMStart[1][1] && addr < NWRAMEnd[1][1])
-        {
-            u8* ptr = NWRAMMap_B[1][(addr >> 15) & NWRAMMask[1][1]];
-            if (ptr)
+            if (addr >= NWRAMStart[1][1] && addr < NWRAMEnd[1][1])
             {
-                *(u8*)&ptr[addr & 0x7FFF] = val;
+                // Write to a bank is special, as it writes to all
+                // parts that are mapped and not just the highest priority
+                // See http://melonds.kuribo64.net/board/thread.php?pid=3974#3974
+                // so we need to iterate through all parts and write to all mapped here
+                unsigned int destPartSetting = ((addr >> 13) & (NWRAMMask[1][1] << 2)) | 0x81;
+                for (int page = 0; page < 8; page++)
+                {
+                    unsigned int bankInfo = (MBK[1][1 + (page / 4)] >> ((page % 4) * 8)) & 0xff;
+                    if (bankInfo != destPartSetting)
+                        continue;
+                    u8* ptr = &NWRAM_B[page * 0x8000];
+                    *(u8*)&ptr[addr & 0x7FFF] = val;
 #ifdef JIT_ENABLED
-                ARMJIT::CheckAndInvalidate<1, ARMJIT_Memory::memregion_NewSharedWRAM_B>(addr);
+                    ARMJIT::CheckAndInvalidate<1, ARMJIT_Memory::memregion_NewSharedWRAM_B>(addr);
 #endif
+                }
+                return;
             }
-            return;
-        }
-        if (addr >= NWRAMStart[1][2] && addr < NWRAMEnd[1][2])
-        {
-            u8* ptr = NWRAMMap_C[1][(addr >> 15) & NWRAMMask[1][2]];
-            if (ptr)
+            if (addr >= NWRAMStart[1][2] && addr < NWRAMEnd[1][2])
             {
-                *(u8*)&ptr[addr & 0x7FFF] = val;
+                // Write to a bank is special, as it writes to all
+                // parts that are mapped and not just the highest priority
+                // See http://melonds.kuribo64.net/board/thread.php?pid=3974#3974
+                // so we need to iterate through all parts and write to all mapped here
+                unsigned int destPartSetting = ((addr >> 13) & (NWRAMMask[1][2] << 2)) | 0x81;
+                for (int page = 0; page < 8; page++)
+                {
+                    unsigned int bankInfo = (MBK[1][3 + (page / 4)] >> ((page % 4) * 8)) & 0xff;
+                    if (bankInfo != destPartSetting)
+                        continue;
+                    u8* ptr = &NWRAM_C[page * 0x8000];
+                    *(u8*)&ptr[addr & 0x7FFF] = val;
 #ifdef JIT_ENABLED
-                ARMJIT::CheckAndInvalidate<1, ARMJIT_Memory::memregion_NewSharedWRAM_C>(addr);
+                    ARMJIT::CheckAndInvalidate<1, ARMJIT_Memory::memregion_NewSharedWRAM_C>(addr);
 #endif
+                }
+                return;
             }
-            return;
         }
         return NDS::ARM7Write8(addr, val);
 
@@ -1351,41 +1581,69 @@ void ARM7Write16(u32 addr, u16 val)
     switch (addr & 0xFF800000)
     {
     case 0x03000000:
-        if (addr >= NWRAMStart[1][0] && addr < NWRAMEnd[1][0])
+    case 0x03800000:
+        if (SCFG_EXT[1] & (1 << 25))
         {
-            u8* ptr = NWRAMMap_A[1][(addr >> 16) & NWRAMMask[1][0]];
-            if (ptr)
+            if (addr >= NWRAMStart[1][0] && addr < NWRAMEnd[1][0])
             {
-                *(u16*)&ptr[addr & 0xFFFF] = val;
+                // Write to a bank is special, as it writes to all
+                // parts that are mapped and not just the highest priority
+                // See http://melonds.kuribo64.net/board/thread.php?pid=3974#3974
+                // so we need to iterate through all parts and write to all mapped here
+                unsigned int destPartSetting = ((addr >> 13) & (NWRAMMask[1][0] << 3)) | 0x81;
+                for (int page = 0; page < 4; page++)
+                {
+                    unsigned int bankInfo = (MBK[1][0 + (page / 4)] >> ((page % 4) * 8)) & 0xff;
+                    if (bankInfo != destPartSetting)
+                        continue;
+                    u8* ptr = &NWRAM_A[page * 0x8000];
+                    *(u16*)&ptr[addr & 0x7FFF] = val;
 #ifdef JIT_ENABLED
-                ARMJIT::CheckAndInvalidate<1, ARMJIT_Memory::memregion_NewSharedWRAM_A>(addr);
+                    ARMJIT::CheckAndInvalidate<1, ARMJIT_Memory::memregion_NewSharedWRAM_A>(addr);
 #endif
+                }
+                return;
             }
-            return;
-        }
-        if (addr >= NWRAMStart[1][1] && addr < NWRAMEnd[1][1])
-        {
-            u8* ptr = NWRAMMap_B[1][(addr >> 15) & NWRAMMask[1][1]];
-            if (ptr)
+            if (addr >= NWRAMStart[1][1] && addr < NWRAMEnd[1][1])
             {
-                *(u16*)&ptr[addr & 0x7FFF] = val;
+                // Write to a bank is special, as it writes to all
+                // parts that are mapped and not just the highest priority
+                // See http://melonds.kuribo64.net/board/thread.php?pid=3974#3974
+                // so we need to iterate through all parts and write to all mapped here
+                unsigned int destPartSetting = ((addr >> 13) & (NWRAMMask[1][1] << 2)) | 0x81;
+                for (int page = 0; page < 8; page++)
+                {
+                    unsigned int bankInfo = (MBK[1][1 + (page / 4)] >> ((page % 4) * 8)) & 0xff;
+                    if (bankInfo != destPartSetting)
+                        continue;
+                    u8* ptr = &NWRAM_B[page * 0x8000];
+                    *(u16*)&ptr[addr & 0x7FFF] = val;
 #ifdef JIT_ENABLED
-                ARMJIT::CheckAndInvalidate<1, ARMJIT_Memory::memregion_NewSharedWRAM_B>(addr);
+                    ARMJIT::CheckAndInvalidate<1, ARMJIT_Memory::memregion_NewSharedWRAM_B>(addr);
 #endif
+                }
+                return;
             }
-            return;
-        }
-        if (addr >= NWRAMStart[1][2] && addr < NWRAMEnd[1][2])
-        {
-            u8* ptr = NWRAMMap_C[1][(addr >> 15) & NWRAMMask[1][2]];
-            if (ptr)
+            if (addr >= NWRAMStart[1][2] && addr < NWRAMEnd[1][2])
             {
-                *(u16*)&ptr[addr & 0x7FFF] = val;
+                // Write to a bank is special, as it writes to all
+                // parts that are mapped and not just the highest priority
+                // See http://melonds.kuribo64.net/board/thread.php?pid=3974#3974
+                // so we need to iterate through all parts and write to all mapped here
+                unsigned int destPartSetting = ((addr >> 13) & (NWRAMMask[1][2] << 2)) | 0x81;
+                for (int page = 0; page < 8; page++)
+                {
+                    unsigned int bankInfo = (MBK[1][3 + (page / 4)] >> ((page % 4) * 8)) & 0xff;
+                    if (bankInfo != destPartSetting)
+                        continue;
+                    u8* ptr = &NWRAM_C[page * 0x8000];
+                    *(u16*)&ptr[addr & 0x7FFF] = val;
 #ifdef JIT_ENABLED
-                ARMJIT::CheckAndInvalidate<1, ARMJIT_Memory::memregion_NewSharedWRAM_C>(addr);
+                    ARMJIT::CheckAndInvalidate<1, ARMJIT_Memory::memregion_NewSharedWRAM_C>(addr);
 #endif
+                }
+                return;
             }
-            return;
         }
         return NDS::ARM7Write16(addr, val);
 
@@ -1410,41 +1668,69 @@ void ARM7Write32(u32 addr, u32 val)
     switch (addr & 0xFF800000)
     {
     case 0x03000000:
-        if (addr >= NWRAMStart[1][0] && addr < NWRAMEnd[1][0])
+    case 0x03800000:
+        if (SCFG_EXT[1] & (1 << 25))
         {
-            u8* ptr = NWRAMMap_A[1][(addr >> 16) & NWRAMMask[1][0]];
-            if (ptr)
+            if (addr >= NWRAMStart[1][0] && addr < NWRAMEnd[1][0])
             {
-                *(u32*)&ptr[addr & 0xFFFF] = val;
+                // Write to a bank is special, as it writes to all
+                // parts that are mapped and not just the highest priority
+                // See http://melonds.kuribo64.net/board/thread.php?pid=3974#3974
+                // so we need to iterate through all parts and write to all mapped here
+                unsigned int destPartSetting = ((addr >> 13) & (NWRAMMask[1][0] << 3)) | 0x81;
+                for (int page = 0; page < 4; page++)
+                {
+                    unsigned int bankInfo = (MBK[1][0 + (page / 4)] >> ((page % 4) * 8)) & 0xff;
+                    if (bankInfo != destPartSetting)
+                        continue;
+                    u8* ptr = &NWRAM_A[page * 0x8000];
+                    *(u32*)&ptr[addr & 0x7FFF] = val;
 #ifdef JIT_ENABLED
-                ARMJIT::CheckAndInvalidate<1, ARMJIT_Memory::memregion_NewSharedWRAM_A>(addr);
+                    ARMJIT::CheckAndInvalidate<1, ARMJIT_Memory::memregion_NewSharedWRAM_A>(addr);
 #endif
+                }
+                return;
             }
-            return;
-        }
-        if (addr >= NWRAMStart[1][1] && addr < NWRAMEnd[1][1])
-        {
-            u8* ptr = NWRAMMap_B[1][(addr >> 15) & NWRAMMask[1][1]];
-            if (ptr)
+            if (addr >= NWRAMStart[1][1] && addr < NWRAMEnd[1][1])
             {
-                *(u32*)&ptr[addr & 0x7FFF] = val;
+                // Write to a bank is special, as it writes to all
+                // parts that are mapped and not just the highest priority
+                // See http://melonds.kuribo64.net/board/thread.php?pid=3974#3974
+                // so we need to iterate through all parts and write to all mapped here
+                unsigned int destPartSetting = ((addr >> 13) & (NWRAMMask[1][1] << 2)) | 0x81;
+                for (int page = 0; page < 8; page++)
+                {
+                    unsigned int bankInfo = (MBK[1][1 + (page / 4)] >> ((page % 4) * 8)) & 0xff;
+                    if (bankInfo != destPartSetting)
+                        continue;
+                    u8* ptr = &NWRAM_B[page * 0x8000];
+                    *(u32*)&ptr[addr & 0x7FFF] = val;
 #ifdef JIT_ENABLED
-                ARMJIT::CheckAndInvalidate<1, ARMJIT_Memory::memregion_NewSharedWRAM_B>(addr);
+                    ARMJIT::CheckAndInvalidate<1, ARMJIT_Memory::memregion_NewSharedWRAM_B>(addr);
 #endif
+                }
+                return;
             }
-            return;
-        }
-        if (addr >= NWRAMStart[1][2] && addr < NWRAMEnd[1][2])
-        {
-            u8* ptr = NWRAMMap_C[1][(addr >> 15) & NWRAMMask[1][2]];
-            if (ptr)
+            if (addr >= NWRAMStart[1][2] && addr < NWRAMEnd[1][2])
             {
-                *(u32*)&ptr[addr & 0x7FFF] = val;
+                // Write to a bank is special, as it writes to all
+                // parts that are mapped and not just the highest priority
+                // See http://melonds.kuribo64.net/board/thread.php?pid=3974#3974
+                // so we need to iterate through all parts and write to all mapped here
+                unsigned int destPartSetting = ((addr >> 13) & (NWRAMMask[1][2] << 2)) | 0x81;
+                for (int page = 0; page < 8; page++)
+                {
+                    unsigned int bankInfo = (MBK[1][3 + (page / 4)] >> ((page % 4) * 8)) & 0xff;
+                    if (bankInfo != destPartSetting)
+                        continue;
+                    u8* ptr = &NWRAM_C[page * 0x8000];
+                    *(u32*)&ptr[addr & 0x7FFF] = val;
 #ifdef JIT_ENABLED
-                ARMJIT::CheckAndInvalidate<1, ARMJIT_Memory::memregion_NewSharedWRAM_C>(addr);
+                    ARMJIT::CheckAndInvalidate<1, ARMJIT_Memory::memregion_NewSharedWRAM_C>(addr);
 #endif
+                }
+                return;
             }
-            return;
         }
         return NDS::ARM7Write32(addr, val);
 
@@ -1643,30 +1929,44 @@ void ARM9IOWrite8(u32 addr, u8 val)
         return;
 
     case 0x04004006:
+        if (!(SCFG_EXT[1] & (1 << 31))) /* no access to SCFG Registers if disabled*/
+            return;
         SCFG_RST = (SCFG_RST & 0xFF00) | val;
         DSi_DSP::SetRstLine(val & 1);
         return;
 
-    case 0x04004040: MapNWRAM_A(0, val); return;
-    case 0x04004041: MapNWRAM_A(1, val); return;
-    case 0x04004042: MapNWRAM_A(2, val); return;
-    case 0x04004043: MapNWRAM_A(3, val); return;
-    case 0x04004044: MapNWRAM_B(0, val); return;
-    case 0x04004045: MapNWRAM_B(1, val); return;
-    case 0x04004046: MapNWRAM_B(2, val); return;
-    case 0x04004047: MapNWRAM_B(3, val); return;
-    case 0x04004048: MapNWRAM_B(4, val); return;
-    case 0x04004049: MapNWRAM_B(5, val); return;
-    case 0x0400404A: MapNWRAM_B(6, val); return;
-    case 0x0400404B: MapNWRAM_B(7, val); return;
-    case 0x0400404C: MapNWRAM_C(0, val); return;
-    case 0x0400404D: MapNWRAM_C(1, val); return;
-    case 0x0400404E: MapNWRAM_C(2, val); return;
-    case 0x0400404F: MapNWRAM_C(3, val); return;
-    case 0x04004050: MapNWRAM_C(4, val); return;
-    case 0x04004051: MapNWRAM_C(5, val); return;
-    case 0x04004052: MapNWRAM_C(6, val); return;
-    case 0x04004053: MapNWRAM_C(7, val); return;
+    case 0x04004040:
+    case 0x04004041:
+    case 0x04004042:
+    case 0x04004043:
+        if (!(SCFG_EXT[1] & (1 << 31))) /* no access to SCFG Registers if disabled*/
+            return;
+        MapNWRAM_A(addr & 3, val);
+        return;
+    case 0x04004044:
+    case 0x04004045:
+    case 0x04004046:
+    case 0x04004047:
+    case 0x04004048:
+    case 0x04004049:
+    case 0x0400404A:
+    case 0x0400404B:
+        if (!(SCFG_EXT[1] & (1 << 31))) /* no access to SCFG Registers if disabled*/
+            return;
+        MapNWRAM_B((addr - 0x04) & 7, val);
+        return;
+    case 0x0400404C:
+    case 0x0400404D:
+    case 0x0400404E:
+    case 0x0400404F:
+    case 0x04004050:
+    case 0x04004051:
+    case 0x04004052:
+    case 0x04004053:
+        if (!(SCFG_EXT[1] & (1 << 31))) /* no access to SCFG Registers if disabled*/
+            return;
+        MapNWRAM_C((addr-0x0C) & 7, val);
+        return;
     }
 
     if ((addr & 0xFFFFFF00) == 0x04004200)
@@ -1689,53 +1989,43 @@ void ARM9IOWrite16(u32 addr, u16 val)
     switch (addr)
     {
     case 0x04004004:
+        if (!(SCFG_EXT[0] & (1 << 31))) /* no access to SCFG Registers if disabled*/
+            return;
         Set_SCFG_Clock9(val);
         return;
 
     case 0x04004006:
+        if (!(SCFG_EXT[0] & (1 << 31))) /* no access to SCFG Registers if disabled*/
+            return;
         SCFG_RST = val;
         DSi_DSP::SetRstLine(val & 1);
         return;
 
     case 0x04004040:
-        MapNWRAM_A(0, val & 0xFF);
-        MapNWRAM_A(1, val >> 8);
-        return;
     case 0x04004042:
-        MapNWRAM_A(2, val & 0xFF);
-        MapNWRAM_A(3, val >> 8);
+        if (!(SCFG_EXT[0] & (1 << 31))) /* no access to SCFG Registers if disabled*/
+            return;
+        MapNWRAM_A((addr & 2), val & 0xFF);
+        MapNWRAM_A((addr & 2) + 1, val >> 8);
         return;
+
     case 0x04004044:
-        MapNWRAM_B(0, val & 0xFF);
-        MapNWRAM_B(1, val >> 8);
-        return;
     case 0x04004046:
-        MapNWRAM_B(2, val & 0xFF);
-        MapNWRAM_B(3, val >> 8);
-        return;
     case 0x04004048:
-        MapNWRAM_B(4, val & 0xFF);
-        MapNWRAM_B(5, val >> 8);
-        return;
     case 0x0400404A:
-        MapNWRAM_B(6, val & 0xFF);
-        MapNWRAM_B(7, val >> 8);
+        if (!(SCFG_EXT[0] & (1 << 31))) /* no access to SCFG Registers if disabled*/
+            return;
+        MapNWRAM_B(((addr - 0x04) & 6), val & 0xFF);
+        MapNWRAM_B(((addr - 0x04) & 6) + 1, val >> 8);
         return;
     case 0x0400404C:
-        MapNWRAM_C(0, val & 0xFF);
-        MapNWRAM_C(1, val >> 8);
-        return;
     case 0x0400404E:
-        MapNWRAM_C(2, val & 0xFF);
-        MapNWRAM_C(3, val >> 8);
-        return;
     case 0x04004050:
-        MapNWRAM_C(4, val & 0xFF);
-        MapNWRAM_C(5, val >> 8);
-        return;
     case 0x04004052:
-        MapNWRAM_C(6, val & 0xFF);
-        MapNWRAM_C(7, val >> 8);
+        if (!(SCFG_EXT[0] & (1 << 31))) /* no access to SCFG Registers if disabled*/
+            return;
+        MapNWRAM_C(((addr - 0x0C) & 6), val & 0xFF);
+        MapNWRAM_C(((addr - 0x0C) & 6) + 1, val >> 8);
         return;
     }
 
@@ -1759,6 +2049,8 @@ void ARM9IOWrite32(u32 addr, u32 val)
     switch (addr)
     {
     case 0x04004004:
+        if (!(SCFG_EXT[0] & (1 << 31))) /* no access to SCFG Registers if disabled*/
+            return;
         Set_SCFG_Clock9(val & 0xFFFF);
         SCFG_RST = val >> 16;
         DSi_DSP::SetRstLine((val >> 16) & 1);
@@ -1766,6 +2058,8 @@ void ARM9IOWrite32(u32 addr, u32 val)
 
     case 0x04004008:
         {
+            if (!(SCFG_EXT[0] & (1 << 31))) /* no access to SCFG Registers if disabled*/
+                return;
             u32 oldram = (SCFG_EXT[0] >> 14) & 0x3;
             u32 newram = (val >> 14) & 0x3;
 
@@ -1799,38 +2093,60 @@ void ARM9IOWrite32(u32 addr, u32 val)
         return;
 
     case 0x04004040:
+        if (!(SCFG_EXT[0] & (1 << 31))) /* no access to SCFG Registers if disabled*/
+            return;
         MapNWRAM_A(0, val & 0xFF);
         MapNWRAM_A(1, (val >> 8) & 0xFF);
         MapNWRAM_A(2, (val >> 16) & 0xFF);
         MapNWRAM_A(3, val >> 24);
         return;
     case 0x04004044:
+        if (!(SCFG_EXT[0] & (1 << 31))) /* no access to SCFG Registers if disabled*/
+            return;
         MapNWRAM_B(0, val & 0xFF);
         MapNWRAM_B(1, (val >> 8) & 0xFF);
         MapNWRAM_B(2, (val >> 16) & 0xFF);
         MapNWRAM_B(3, val >> 24);
         return;
     case 0x04004048:
+        if (!(SCFG_EXT[0] & (1 << 31))) /* no access to SCFG Registers if disabled*/
+            return;
         MapNWRAM_B(4, val & 0xFF);
         MapNWRAM_B(5, (val >> 8) & 0xFF);
         MapNWRAM_B(6, (val >> 16) & 0xFF);
         MapNWRAM_B(7, val >> 24);
         return;
     case 0x0400404C:
+        if (!(SCFG_EXT[0] & (1 << 31))) /* no access to SCFG Registers if disabled*/
+            return;
         MapNWRAM_C(0, val & 0xFF);
         MapNWRAM_C(1, (val >> 8) & 0xFF);
         MapNWRAM_C(2, (val >> 16) & 0xFF);
         MapNWRAM_C(3, val >> 24);
         return;
     case 0x04004050:
+        if (!(SCFG_EXT[0] & (1 << 31))) /* no access to SCFG Registers if disabled*/
+            return;
         MapNWRAM_C(4, val & 0xFF);
         MapNWRAM_C(5, (val >> 8) & 0xFF);
         MapNWRAM_C(6, (val >> 16) & 0xFF);
         MapNWRAM_C(7, val >> 24);
         return;
-    case 0x04004054: MapNWRAMRange(0, 0, val); return;
-    case 0x04004058: MapNWRAMRange(0, 1, val); return;
-    case 0x0400405C: MapNWRAMRange(0, 2, val); return;
+    case 0x04004054: 
+        if (!(SCFG_EXT[0] & (1 << 31))) /* no access to SCFG Registers if disabled*/
+            return;
+        MapNWRAMRange(0, 0, val);
+        return;
+    case 0x04004058: 
+        if (!(SCFG_EXT[0] & (1 << 31))) /* no access to SCFG Registers if disabled*/
+            return;
+        MapNWRAMRange(0, 1, val); 
+        return;
+    case 0x0400405C: 
+        if (!(SCFG_EXT[0] & (1 << 31))) /* no access to SCFG Registers if disabled*/
+            return;
+        MapNWRAMRange(0, 2, val);
+        return;
 
     case 0x04004100: NDMACnt[0] = val & 0x800F0000; return;
     case 0x04004104: NDMAs[0]->SrcAddr = val & 0xFFFFFFFC; return;
@@ -1877,7 +2193,8 @@ u8 ARM7IORead8(u32 addr)
 {
     switch (addr)
     {
-    case 0x04004000: return SCFG_BIOS & 0xFF;
+    case 0x04004000: 
+        return SCFG_BIOS & 0xFF;
     case 0x04004001: return SCFG_BIOS >> 8;
 
     CASE_READ8_32BIT(0x04004040, MBK[1][0])
@@ -2026,11 +2343,29 @@ void ARM7IOWrite8(u32 addr, u8 val)
     switch (addr)
     {
     case 0x04004000:
+        if (!(SCFG_EXT[1] & (1 << 31))) /* no access to SCFG Registers if disabled*/
+            return;
         SCFG_BIOS |= (val & 0x03);
         return;
     case 0x04004001:
+        if (!(SCFG_EXT[1] & (1 << 31))) /* no access to SCFG Registers if disabled*/
+            return;
         SCFG_BIOS |= ((val & 0x07) << 8);
         return;
+    case 0x04004060:
+    case 0x04004061:
+    case 0x04004062:
+    case 0x04004063:
+    {
+        if (!(SCFG_EXT[1] & (1 << 31))) /* no access to SCFG Registers if disabled*/
+            return;
+        u32 tmp = MBK[0][8];
+        tmp &= ~(0xff << ((addr % 4) * 8));
+        tmp |= (val << ((addr % 4) * 8));
+        MBK[0][8] = tmp & 0x00FFFF0F;
+        MBK[1][8] = MBK[0][8];
+        return;
+    }
 
     case 0x04004500: DSi_I2C::WriteData(val); return;
     case 0x04004501: DSi_I2C::WriteCnt(val); return;
@@ -2043,18 +2378,34 @@ void ARM7IOWrite16(u32 addr, u16 val)
 {
     switch (addr)
     {
-    case 0x04000218: NDS::IE2 = (val & 0x7FF7); NDS::UpdateIRQ(1); return;
-    case 0x0400021C: NDS::IF2 &= ~(val & 0x7FF7); NDS::UpdateIRQ(1); return;
+        case 0x04000218: NDS::IE2 = (val & 0x7FF7); NDS::UpdateIRQ(1); return;
+        case 0x0400021C: NDS::IF2 &= ~(val & 0x7FF7); NDS::UpdateIRQ(1); return;
 
-    case 0x04004000:
-        SCFG_BIOS |= (val & 0x0703);
-        return;
-    case 0x04004004:
-        SCFG_Clock7 = val & 0x0187;
-        return;
-    case 0x04004010:
-        Set_SCFG_MC((SCFG_MC & 0xFFFF0000) | val);
-        return;
+        case 0x04004000:
+            if (!(SCFG_EXT[1] & (1 << 31))) /* no access to SCFG Registers if disabled*/
+                return;
+            SCFG_BIOS |= (val & 0x0703);
+            return;
+        case 0x04004004:
+            if (!(SCFG_EXT[1] & (1 << 31))) /* no access to SCFG Registers if disabled*/
+                return;
+            SCFG_Clock7 = val & 0x0187;
+            return;
+        case 0x04004010:
+            if (!(SCFG_EXT[1] & (1 << 31))) /* no access to SCFG Registers if disabled*/
+                return;
+            Set_SCFG_MC((SCFG_MC & 0xFFFF0000) | val);
+            return;
+        case 0x04004060:
+        case 0x04004062:
+            if (!(SCFG_EXT[1] & (1 << 31))) /* no access to SCFG Registers if disabled*/
+                return;
+            u32 tmp = MBK[0][8];
+            tmp &= ~(0xffff << ((addr % 4) * 8));
+            tmp |= (val << ((addr % 4) * 8));
+            MBK[0][8] = tmp & 0x00FFFF0F;
+            MBK[1][8] = MBK[0][8];
+            return;
     }
 
     if (addr >= 0x04004800 && addr < 0x04004A00)
@@ -2079,9 +2430,13 @@ void ARM7IOWrite32(u32 addr, u32 val)
     case 0x0400021C: NDS::IF2 &= ~(val & 0x7FF7); NDS::UpdateIRQ(1); return;
 
     case 0x04004000:
+        if (!(SCFG_EXT[1] & (1 << 31))) /* no access to SCFG Registers if disabled*/
+            return;
         SCFG_BIOS |= (val & 0x0703);
         return;
     case 0x04004008:
+        if (!(SCFG_EXT[1] & (1 << 31))) /* no access to SCFG Registers if disabled*/
+            return;
         SCFG_EXT[0] &= ~0x03000000;
         SCFG_EXT[0] |= (val & 0x03000000);
         SCFG_EXT[1] &= ~0x93FF0F07;
@@ -2089,13 +2444,33 @@ void ARM7IOWrite32(u32 addr, u32 val)
         printf("SCFG_EXT = %08X / %08X (val7 %08X)\n", SCFG_EXT[0], SCFG_EXT[1], val);
         return;
     case 0x04004010:
+        if (!(SCFG_EXT[1] & (1 << 31))) /* no access to SCFG Registers if disabled*/
+            return;
         Set_SCFG_MC(val);
         return;
 
-    case 0x04004054: MapNWRAMRange(1, 0, val); return;
-    case 0x04004058: MapNWRAMRange(1, 1, val); return;
-    case 0x0400405C: MapNWRAMRange(1, 2, val); return;
-    case 0x04004060: val &= 0x00FFFF0F; MBK[0][8] = val; MBK[1][8] = val; return;
+    case 0x04004054: 
+        if (!(SCFG_EXT[1] & (1 << 31))) /* no access to SCFG Registers if disabled*/
+            return;
+        MapNWRAMRange(1, 0, val);
+        return;
+    case 0x04004058: 
+        if (!(SCFG_EXT[1] & (1 << 31))) /* no access to SCFG Registers if disabled*/
+            return;
+        MapNWRAMRange(1, 1, val);
+        return;
+    case 0x0400405C: 
+        if (!(SCFG_EXT[1] & (1 << 31))) /* no access to SCFG Registers if disabled*/
+            return;
+        MapNWRAMRange(1, 2, val);
+        return;
+    case 0x04004060: 
+        if (!(SCFG_EXT[1] & (1 << 31))) /* no access to SCFG Registers if disabled*/
+            return;
+        val &= 0x00FFFF0F;
+        MBK[0][8] = val; 
+        MBK[1][8] = val; 
+        return;
 
     case 0x04004100: NDMACnt[1] = val & 0x800F0000; return;
     case 0x04004104: NDMAs[4]->SrcAddr = val & 0xFFFFFFFC; return;
diff --git a/src/DSi.h b/src/DSi.h
index a56df66..187dc52 100644
--- a/src/DSi.h
+++ b/src/DSi.h
@@ -60,6 +60,7 @@ bool Init();
 void DeInit();
 void Reset();
 
+void SetupDirectBoot();
 void SoftReset();
 
 bool LoadBIOS();
diff --git a/src/NDS.cpp b/src/NDS.cpp
index 699b3a6..cd0cb92 100644
--- a/src/NDS.cpp
+++ b/src/NDS.cpp
@@ -164,7 +164,6 @@ bool Running;
 
 bool RunningGame;
 
-
 void DivDone(u32 param);
 void SqrtDone(u32 param);
 void RunTimer(u32 tid, s32 cycles);
@@ -342,8 +341,13 @@ void SetupDirectBoot()
 {
     if (ConsoleType == 1)
     {
-        printf("!! DIRECT BOOT NOT SUPPORTED IN DSI MODE\n");
-        return;
+        // With the BIOS select in SCFG_BIOS and the initialization od
+        // SCFG_BIOS depending on the Header->UnitType, we can now boot 
+        // directly in the roms.
+        // There are some more SCFG Settings that change depending on
+        // the unit type, so this is experimental
+        printf("!! DIRECT BOOT NOT STABLE IN DSI MODE\n");
+        DSi::SetupDirectBoot();
     }
 
     u32 bootparams[8];