diff --git a/ARMInterpreter_LoadStore.cpp b/ARMInterpreter_LoadStore.cpp
index 583f175d..4a78186c 100644
--- a/ARMInterpreter_LoadStore.cpp
+++ b/ARMInterpreter_LoadStore.cpp
@@ -477,6 +477,39 @@ s32 T_LDRB_REG(ARM* cpu)
 }
 
 
+s32 T_STRH_REG(ARM* cpu)
+{
+    u32 addr = cpu->R[(cpu->CurInstr >> 3) & 0x7] + cpu->R[(cpu->CurInstr >> 6) & 0x7];
+    cpu->Write16(addr, cpu->R[cpu->CurInstr & 0x7]);
+
+    return C_N(2) + cpu->MemWaitstate(2, addr);
+}
+
+s32 T_LDRSB_REG(ARM* cpu)
+{
+    u32 addr = cpu->R[(cpu->CurInstr >> 3) & 0x7] + cpu->R[(cpu->CurInstr >> 6) & 0x7];
+    cpu->R[cpu->CurInstr & 0x7] = (s32)(s8)cpu->Read8(addr);
+
+    return C_S(1) + C_N(1) + C_I(1) + cpu->MemWaitstate(3, addr);
+}
+
+s32 T_LDRH_REG(ARM* cpu)
+{
+    u32 addr = cpu->R[(cpu->CurInstr >> 3) & 0x7] + cpu->R[(cpu->CurInstr >> 6) & 0x7];
+    cpu->R[cpu->CurInstr & 0x7] = cpu->Read16(addr);
+
+    return C_S(1) + C_N(1) + C_I(1) + cpu->MemWaitstate(2, addr);
+}
+
+s32 T_LDRSH_REG(ARM* cpu)
+{
+    u32 addr = cpu->R[(cpu->CurInstr >> 3) & 0x7] + cpu->R[(cpu->CurInstr >> 6) & 0x7];
+    cpu->R[cpu->CurInstr & 0x7] = (s32)(s16)cpu->Read16(addr);
+
+    return C_S(1) + C_N(1) + C_I(1) + cpu->MemWaitstate(2, addr);
+}
+
+
 s32 T_STR_IMM(ARM* cpu)
 {
     u32 offset = (cpu->CurInstr >> 4) & 0x7C;
@@ -620,6 +653,48 @@ s32 T_POP(ARM* cpu)
     return cycles;
 }
 
+s32 T_STMIA(ARM* cpu)
+{
+    u32 base = cpu->R[(cpu->CurInstr >> 8) & 0x7];
+
+    int cycles = C_N(2);
+
+    for (int i = 0; i < 8; i++)
+    {
+        if (cpu->CurInstr & (1<<i))
+        {
+            cpu->Write32(base, cpu->R[i]);
+            cycles += C_S(1) + cpu->MemWaitstate(3, base);
+            base += 4;
+        }
+    }
+
+    cpu->R[(cpu->CurInstr >> 8) & 0x7] = base;
+
+    return cycles - C_S(1);
+}
+
+s32 T_LDMIA(ARM* cpu)
+{
+    u32 base = cpu->R[(cpu->CurInstr >> 8) & 0x7];
+
+    int cycles = C_N(1) + C_I(1);
+
+    for (int i = 0; i < 8; i++)
+    {
+        if (cpu->CurInstr & (1<<i))
+        {
+            cpu->R[i] = cpu->Read32(base);
+            cycles += C_S(1) + cpu->MemWaitstate(3, base);
+            base += 4;
+        }
+    }
+
+    cpu->R[(cpu->CurInstr >> 8) & 0x7] = base;
+
+    return cycles;
+}
+
 
 }
 
diff --git a/ARMInterpreter_LoadStore.h b/ARMInterpreter_LoadStore.h
index 61b1a39f..4759d1f5 100644
--- a/ARMInterpreter_LoadStore.h
+++ b/ARMInterpreter_LoadStore.h
@@ -48,6 +48,11 @@ s32 T_STRB_REG(ARM* cpu);
 s32 T_LDR_REG(ARM* cpu);
 s32 T_LDRB_REG(ARM* cpu);
 
+s32 T_STRH_REG(ARM* cpu);
+s32 T_LDRSB_REG(ARM* cpu);
+s32 T_LDRH_REG(ARM* cpu);
+s32 T_LDRSH_REG(ARM* cpu);
+
 s32 T_STR_IMM(ARM* cpu);
 s32 T_LDR_IMM(ARM* cpu);
 s32 T_STRB_IMM(ARM* cpu);
@@ -61,6 +66,8 @@ s32 T_LDR_SPREL(ARM* cpu);
 
 s32 T_PUSH(ARM* cpu);
 s32 T_POP(ARM* cpu);
+s32 T_STMIA(ARM* cpu);
+s32 T_LDMIA(ARM* cpu);
 
 }
 
diff --git a/ARM_InstrTable.h b/ARM_InstrTable.h
index 78dd18c2..03f7c3a7 100644
--- a/ARM_InstrTable.h
+++ b/ARM_InstrTable.h
@@ -1695,26 +1695,26 @@ INSTRFUNC_PROTO(THUMBInstrTable[1024]) =
     // 0101 0000 00
     T_STR_REG, T_STR_REG, T_STR_REG, T_STR_REG,
     T_STR_REG, T_STR_REG, T_STR_REG, T_STR_REG,
-    T_UNK, T_UNK, T_UNK, T_UNK,
-    T_UNK, T_UNK, T_UNK, T_UNK,
+    T_STRH_REG, T_STRH_REG, T_STRH_REG, T_STRH_REG,
+    T_STRH_REG, T_STRH_REG, T_STRH_REG, T_STRH_REG,
 
     // 0101 0100 00
     T_STRB_REG, T_STRB_REG, T_STRB_REG, T_STRB_REG,
     T_STRB_REG, T_STRB_REG, T_STRB_REG, T_STRB_REG,
-    T_UNK, T_UNK, T_UNK, T_UNK,
-    T_UNK, T_UNK, T_UNK, T_UNK,
+    T_LDRSB_REG, T_LDRSB_REG, T_LDRSB_REG, T_LDRSB_REG,
+    T_LDRSB_REG, T_LDRSB_REG, T_LDRSB_REG, T_LDRSB_REG,
 
     // 0101 1000 00
     T_LDR_REG, T_LDR_REG, T_LDR_REG, T_LDR_REG,
     T_LDR_REG, T_LDR_REG, T_LDR_REG, T_LDR_REG,
-    T_UNK, T_UNK, T_UNK, T_UNK,
-    T_UNK, T_UNK, T_UNK, T_UNK,
+    T_LDRH_REG, T_LDRH_REG, T_LDRH_REG, T_LDRH_REG,
+    T_LDRH_REG, T_LDRH_REG, T_LDRH_REG, T_LDRH_REG,
 
     // 0101 1100 00
     T_LDRB_REG, T_LDRB_REG, T_LDRB_REG, T_LDRB_REG,
     T_LDRB_REG, T_LDRB_REG, T_LDRB_REG, T_LDRB_REG,
-    T_UNK, T_UNK, T_UNK, T_UNK,
-    T_UNK, T_UNK, T_UNK, T_UNK,
+    T_LDRSH_REG, T_LDRSH_REG, T_LDRSH_REG, T_LDRSH_REG,
+    T_LDRSH_REG, T_LDRSH_REG, T_LDRSH_REG, T_LDRSH_REG,
 
     // 0110 0000 00
     T_STR_IMM, T_STR_IMM, T_STR_IMM, T_STR_IMM,
@@ -1865,28 +1865,28 @@ INSTRFUNC_PROTO(THUMBInstrTable[1024]) =
 
 
     // 1100 0000 00
-    T_UNK, T_UNK, T_UNK, T_UNK,
-    T_UNK, T_UNK, T_UNK, T_UNK,
-    T_UNK, T_UNK, T_UNK, T_UNK,
-    T_UNK, T_UNK, T_UNK, T_UNK,
+    T_STMIA, T_STMIA, T_STMIA, T_STMIA,
+    T_STMIA, T_STMIA, T_STMIA, T_STMIA,
+    T_STMIA, T_STMIA, T_STMIA, T_STMIA,
+    T_STMIA, T_STMIA, T_STMIA, T_STMIA,
 
     // 1100 0100 00
-    T_UNK, T_UNK, T_UNK, T_UNK,
-    T_UNK, T_UNK, T_UNK, T_UNK,
-    T_UNK, T_UNK, T_UNK, T_UNK,
-    T_UNK, T_UNK, T_UNK, T_UNK,
+    T_STMIA, T_STMIA, T_STMIA, T_STMIA,
+    T_STMIA, T_STMIA, T_STMIA, T_STMIA,
+    T_STMIA, T_STMIA, T_STMIA, T_STMIA,
+    T_STMIA, T_STMIA, T_STMIA, T_STMIA,
 
     // 1100 1000 00
-    T_UNK, T_UNK, T_UNK, T_UNK,
-    T_UNK, T_UNK, T_UNK, T_UNK,
-    T_UNK, T_UNK, T_UNK, T_UNK,
-    T_UNK, T_UNK, T_UNK, T_UNK,
+    T_LDMIA, T_LDMIA, T_LDMIA, T_LDMIA,
+    T_LDMIA, T_LDMIA, T_LDMIA, T_LDMIA,
+    T_LDMIA, T_LDMIA, T_LDMIA, T_LDMIA,
+    T_LDMIA, T_LDMIA, T_LDMIA, T_LDMIA,
 
     // 1100 1100 00
-    T_UNK, T_UNK, T_UNK, T_UNK,
-    T_UNK, T_UNK, T_UNK, T_UNK,
-    T_UNK, T_UNK, T_UNK, T_UNK,
-    T_UNK, T_UNK, T_UNK, T_UNK,
+    T_LDMIA, T_LDMIA, T_LDMIA, T_LDMIA,
+    T_LDMIA, T_LDMIA, T_LDMIA, T_LDMIA,
+    T_LDMIA, T_LDMIA, T_LDMIA, T_LDMIA,
+    T_LDMIA, T_LDMIA, T_LDMIA, T_LDMIA,
 
     // 1101 0000 00
     T_BCOND, T_BCOND, T_BCOND, T_BCOND,
diff --git a/NDS.cpp b/NDS.cpp
index 23a781a7..1b33f3ed 100644
--- a/NDS.cpp
+++ b/NDS.cpp
@@ -18,6 +18,13 @@ u8 ARM7BIOS[0x4000];
 
 u8 MainRAM[0x400000];
 
+u8 SharedWRAM[0x8000];
+u8 WRAMCnt;
+u8* SWRAM_ARM9;
+u8* SWRAM_ARM7;
+u32 SWRAM_ARM9Mask;
+u32 SWRAM_ARM7Mask;
+
 u8 ARM7WRAM[0x10000];
 
 u8 ARM9ITCM[0x8000];
@@ -68,10 +75,14 @@ void Reset()
     }
 
     memset(MainRAM, 0, 0x400000);
+    memset(SharedWRAM, 0, 0x8000);
     memset(ARM7WRAM, 0, 0x10000);
     memset(ARM9ITCM, 0, 0x8000);
     memset(ARM9DTCM, 0, 0x4000);
 
+    WRAMCnt = 0;
+    MapSharedWRAM();
+
     ARM9ITCMSize = 0;
     ARM9DTCMBase = 0xFFFFFFFF;
     ARM9DTCMSize = 0;
@@ -112,6 +123,41 @@ void Halt()
 }
 
 
+void MapSharedWRAM()
+{
+    switch (WRAMCnt & 0x3)
+    {
+    case 0:
+        SWRAM_ARM9 = &SharedWRAM[0];
+        SWRAM_ARM9Mask = 0x7FFF;
+        SWRAM_ARM7 = NULL;
+        SWRAM_ARM7Mask = 0;
+        break;
+
+    case 1:
+        SWRAM_ARM9 = &SharedWRAM[0x4000];
+        SWRAM_ARM9Mask = 0x3FFF;
+        SWRAM_ARM7 = &SharedWRAM[0];
+        SWRAM_ARM7Mask = 0x3FFF;
+        break;
+
+    case 2:
+        SWRAM_ARM9 = &SharedWRAM[0];
+        SWRAM_ARM9Mask = 0x3FFF;
+        SWRAM_ARM7 = &SharedWRAM[0x4000];
+        SWRAM_ARM7Mask = 0x3FFF;
+        break;
+
+    case 3:
+        SWRAM_ARM9 = NULL;
+        SWRAM_ARM9Mask = 0;
+        SWRAM_ARM7 = &SharedWRAM[0];
+        SWRAM_ARM7Mask = 0x7FFF;
+        break;
+    }
+}
+
+
 
 u8 ARM9Read8(u32 addr)
 {
@@ -132,6 +178,17 @@ u8 ARM9Read8(u32 addr)
     {
     case 0x02000000:
         return *(u8*)&MainRAM[addr & 0x3FFFFF];
+
+    case 0x03000000:
+        if (SWRAM_ARM9) return *(u8*)&SWRAM_ARM9[addr & SWRAM_ARM9Mask];
+        else return 0;
+
+    case 0x04000000:
+        switch (addr)
+        {
+        case 0x04000247:
+            return WRAMCnt;
+        }
     }
 
     printf("unknown arm9 read8 %08X\n", addr);
@@ -158,6 +215,10 @@ u16 ARM9Read16(u32 addr)
     case 0x02000000:
         return *(u16*)&MainRAM[addr & 0x3FFFFF];
 
+    case 0x03000000:
+        if (SWRAM_ARM9) return *(u16*)&SWRAM_ARM9[addr & SWRAM_ARM9Mask];
+        else return 0;
+
     case 0x04000000:
         switch (addr)
         {
@@ -188,6 +249,10 @@ u32 ARM9Read32(u32 addr)
     {
     case 0x02000000:
         return *(u32*)&MainRAM[addr & 0x3FFFFF];
+
+    case 0x03000000:
+        if (SWRAM_ARM9) return *(u32*)&SWRAM_ARM9[addr & SWRAM_ARM9Mask];
+        else return 0;
     }
 
     printf("unknown arm9 read32 %08X | %08X\n", addr, ARM9->R[15]);
@@ -212,6 +277,19 @@ void ARM9Write8(u32 addr, u8 val)
     case 0x02000000:
         *(u8*)&MainRAM[addr & 0x3FFFFF] = val;
         return;
+
+    case 0x03000000:
+        if (SWRAM_ARM9) *(u8*)&SWRAM_ARM9[addr & SWRAM_ARM9Mask] = val;
+        return;
+
+    case 0x04000000:
+        switch (addr)
+        {
+        case 0x04000247:
+            WRAMCnt = val;
+            MapSharedWRAM();
+            return;
+        }
     }
 
     printf("unknown arm9 write8 %08X %02X\n", addr, val);
@@ -236,6 +314,10 @@ void ARM9Write16(u32 addr, u16 val)
         *(u16*)&MainRAM[addr & 0x3FFFFF] = val;
         return;
 
+    case 0x03000000:
+        if (SWRAM_ARM9) *(u16*)&SWRAM_ARM9[addr & SWRAM_ARM9Mask] = val;
+        return;
+
     case 0x04000000:
         switch (addr)
         {
@@ -273,6 +355,10 @@ void ARM9Write32(u32 addr, u32 val)
     case 0x02000000:
         *(u32*)&MainRAM[addr & 0x3FFFFF] = val;
         return;
+
+    case 0x03000000:
+        if (SWRAM_ARM9) *(u32*)&SWRAM_ARM9[addr & SWRAM_ARM9Mask] = val;
+        return;
     }
 
     printf("unknown arm9 write32 %08X %08X | %08X\n", addr, val, ARM9->R[15]);
@@ -292,8 +378,19 @@ u8 ARM7Read8(u32 addr)
     case 0x02000000:
         return *(u8*)&MainRAM[addr & 0x3FFFFF];
 
+    case 0x03000000:
+        if (SWRAM_ARM7) return *(u8*)&SWRAM_ARM7[addr & SWRAM_ARM7Mask];
+        else return *(u8*)&ARM7WRAM[addr & 0xFFFF];
+
     case 0x03800000:
         return *(u8*)&ARM7WRAM[addr & 0xFFFF];
+
+    case 0x04000000:
+        switch (addr)
+        {
+        case 0x04000241:
+            return WRAMCnt;
+        }
     }
 
     printf("unknown arm7 read8 %08X\n", addr);
@@ -312,6 +409,10 @@ u16 ARM7Read16(u32 addr)
     case 0x02000000:
         return *(u16*)&MainRAM[addr & 0x3FFFFF];
 
+    case 0x03000000:
+        if (SWRAM_ARM7) return *(u16*)&SWRAM_ARM7[addr & SWRAM_ARM7Mask];
+        else return *(u16*)&ARM7WRAM[addr & 0xFFFF];
+
     case 0x03800000:
         return *(u16*)&ARM7WRAM[addr & 0xFFFF];
 
@@ -338,8 +439,19 @@ u32 ARM7Read32(u32 addr)
     case 0x02000000:
         return *(u32*)&MainRAM[addr & 0x3FFFFF];
 
+    case 0x03000000:
+        if (SWRAM_ARM7) return *(u32*)&SWRAM_ARM7[addr & SWRAM_ARM7Mask];
+        else return *(u32*)&ARM7WRAM[addr & 0xFFFF];
+
     case 0x03800000:
         return *(u32*)&ARM7WRAM[addr & 0xFFFF];
+
+    case 0x04000000:
+        switch (addr)
+        {
+        case 0x040001A4:
+            return 0x00800000; // hax
+        }
     }
 if ((addr&0xFF000000) == 0xEA000000) Halt();
     printf("unknown arm7 read32 %08X | %08X\n", addr, ARM7->R[15]);
@@ -354,12 +466,17 @@ void ARM7Write8(u32 addr, u8 val)
         *(u8*)&MainRAM[addr & 0x3FFFFF] = val;
         return;
 
+    case 0x03000000:
+        if (SWRAM_ARM7) *(u8*)&SWRAM_ARM7[addr & SWRAM_ARM7Mask] = val;
+        else *(u8*)&ARM7WRAM[addr & 0xFFFF] = val;
+        return;
+
     case 0x03800000:
         *(u8*)&ARM7WRAM[addr & 0xFFFF] = val;
         return;
     }
 
-    printf("unknown arm7 write8 %08X %02X\n", addr, val);
+    printf("unknown arm7 write8 %08X %02X | %08X\n", addr, val, ARM7->R[15]);
 }
 
 void ARM7Write16(u32 addr, u16 val)
@@ -370,6 +487,11 @@ void ARM7Write16(u32 addr, u16 val)
         *(u16*)&MainRAM[addr & 0x3FFFFF] = val;
         return;
 
+    case 0x03000000:
+        if (SWRAM_ARM7) *(u16*)&SWRAM_ARM7[addr & SWRAM_ARM7Mask] = val;
+        else *(u16*)&ARM7WRAM[addr & 0xFFFF] = val;
+        return;
+
     case 0x03800000:
         *(u16*)&ARM7WRAM[addr & 0xFFFF] = val;
         return;
@@ -390,7 +512,7 @@ void ARM7Write16(u32 addr, u16 val)
         }
     }
 
-    printf("unknown arm7 write16 %08X %04X\n", addr, val);
+    printf("unknown arm7 write16 %08X %04X | %08X\n", addr, val, ARM7->R[15]);
 }
 
 void ARM7Write32(u32 addr, u32 val)
@@ -401,12 +523,17 @@ void ARM7Write32(u32 addr, u32 val)
         *(u32*)&MainRAM[addr & 0x3FFFFF] = val;
         return;
 
+    case 0x03000000:
+        if (SWRAM_ARM7) *(u32*)&SWRAM_ARM7[addr & SWRAM_ARM7Mask] = val;
+        else *(u32*)&ARM7WRAM[addr & 0xFFFF] = val;
+        return;
+
     case 0x03800000:
         *(u32*)&ARM7WRAM[addr & 0xFFFF] = val;
         return;
     }
 
-    printf("unknown arm7 write32 %08X %08X\n", addr, val);
+    printf("unknown arm7 write32 %08X %08X | %08X\n", addr, val, ARM7->R[15]);
 }
 
 }
diff --git a/NDS.h b/NDS.h
index 06490425..b0a3311b 100644
--- a/NDS.h
+++ b/NDS.h
@@ -17,6 +17,8 @@ void RunFrame();
 
 void Halt();
 
+void MapSharedWRAM();
+
 u8 ARM9Read8(u32 addr);
 u16 ARM9Read16(u32 addr);
 u32 ARM9Read32(u32 addr);
diff --git a/melonDS.depend b/melonDS.depend
index e36f7dd6..100795b8 100644
--- a/melonDS.depend
+++ b/melonDS.depend
@@ -3,12 +3,12 @@
 	<stdio.h>
 	"NDS.h"
 
-1480776639 c:\documents\sources\melonds\nds.h
+1480786007 c:\documents\sources\melonds\nds.h
 	"types.h"
 
 1463409689 c:\documents\sources\melonds\types.h
 
-1480783776 source:c:\documents\sources\melonds\nds.cpp
+1480784993 source:c:\documents\sources\melonds\nds.cpp
 	<stdio.h>
 	<string.h>
 	"NDS.h"
@@ -25,7 +25,7 @@
 	"types.h"
 	"NDS.h"
 
-1480784235 c:\documents\sources\melonds\arm_instrtable.h
+1480785229 c:\documents\sources\melonds\arm_instrtable.h
 
 1480725698 c:\documents\sources\melonds\arminterpreter.h
 	"types.h"
@@ -49,12 +49,12 @@
 
 1480784137 c:\documents\sources\melonds\arminterpreter_alu.h
 
-1480783339 source:c:\documents\sources\melonds\arminterpreter_alu.cpp
+1480784110 source:c:\documents\sources\melonds\arminterpreter_alu.cpp
 	"ARM.h"
 
-1480780717 c:\documents\sources\melonds\arminterpreter_loadstore.h
+1480785175 c:\documents\sources\melonds\arminterpreter_loadstore.h
 
-1480783030 source:c:\documents\sources\melonds\arminterpreter_loadstore.cpp
+1480785158 source:c:\documents\sources\melonds\arminterpreter_loadstore.cpp
 	<stdio.h>
 	"ARM.h"