Merge pull request #43 from jamsilva/master

SA1 and SuperFX improvements

10 files changed, 201 insertions, 250 deletions

diff --git a/source/cheats.c b/source/cheats.c
index 572fd55..0bedfd4 100644
--- a/source/cheats.c
+++ b/source/cheats.c
@@ -64,7 +64,7 @@ const char* S9xGameGenieToRaw(const char* code, uint32_t* address, uint8_t* byte
    uint32_t data = 0;
    int32_t i;
 
-   if (strlen(code) != 9 || *(code + 4) != '-' || !S9xAllHex(code, 4) || !S9xAllHex(code + 5, 4))
+   if (strlen(code) != 9 || code[4] != '-' || !S9xAllHex(code, 4) || !S9xAllHex(code + 5, 4))
       return "Invalid Game Genie(tm) code - should be 'xxxx-xxxx'.";
 
    strcpy(new_code, "0x");
diff --git a/source/cheats2.c b/source/cheats2.c
index cc303b8..067a929 100644
--- a/source/cheats2.c
+++ b/source/cheats2.c
@@ -81,7 +81,7 @@ void S9xRemoveCheat(uint32_t which1)
       uint8_t* ptr = Memory.Map [block];
 
       if (ptr >= (uint8_t*) MAP_LAST)
-         *(ptr + (address & 0xffff)) = Cheat.c [which1].saved_byte;
+         ptr[address & 0xffff] = Cheat.c [which1].saved_byte;
       else
          S9xSetByte(Cheat.c [which1].saved_byte, address);
       /* Unsave the address for the next call to S9xRemoveCheat. */
@@ -102,7 +102,7 @@ void S9xApplyCheat(uint32_t which1)
    ptr   = Memory.Map [block];
 
    if (ptr >= (uint8_t*) MAP_LAST)
-      *(ptr + (address & 0xffff)) = Cheat.c [which1].byte;
+      ptr[address & 0xffff] = Cheat.c [which1].byte;
    else
       S9xSetByte(Cheat.c [which1].byte, address);
    Cheat.c [which1].saved = true;
diff --git a/source/cpuaddr.h b/source/cpuaddr.h
index bdeb242..2d638b9 100644
--- a/source/cpuaddr.h
+++ b/source/cpuaddr.h
@@ -33,7 +33,7 @@ static INLINE void RelativeLong(void)
 #ifdef FAST_LSB_WORD_ACCESS
    OpAddress = *(uint16_t*) CPU.PC;
 #else
-   OpAddress = *CPU.PC + (*(CPU.PC + 1) << 8);
+   OpAddress = CPU.PC[0] + (CPU.PC[1] << 8);
 #endif
 #ifndef SA1_OPCODES
    CPU.Cycles += CPU.MemSpeedx2 + ONE_CYCLE;
@@ -48,12 +48,12 @@ static INLINE void AbsoluteIndexedIndirect(bool read)
 #ifdef FAST_LSB_WORD_ACCESS
    OpAddress = (ICPU.Registers.X.W + * (uint16_t*) CPU.PC) & 0xffff;
 #else
-   OpAddress = (ICPU.Registers.X.W + *CPU.PC + (*(CPU.PC + 1) << 8)) & 0xffff;
+   OpAddress = (ICPU.Registers.X.W + CPU.PC[0] + (CPU.PC[1] << 8)) & 0xffff;
 #endif
 #ifndef SA1_OPCODES
    CPU.Cycles += CPU.MemSpeedx2;
 #endif
-   OpenBus = *(CPU.PC + 1);
+   OpenBus = CPU.PC[1];
    CPU.PC += 2;
    OpAddress = S9xGetWord(ICPU.ShiftedPB + OpAddress);
    if (read)
@@ -65,12 +65,12 @@ static INLINE void AbsoluteIndirectLong(bool read)
 #ifdef FAST_LSB_WORD_ACCESS
    OpAddress = *(uint16_t*) CPU.PC;
 #else
-   OpAddress = *CPU.PC + (*(CPU.PC + 1) << 8);
+   OpAddress = CPU.PC[0] + (CPU.PC[1] << 8);
 #endif
 #ifndef SA1_OPCODES
    CPU.Cycles += CPU.MemSpeedx2;
 #endif
-   OpenBus = *(CPU.PC + 1);
+   OpenBus = CPU.PC[1];
    CPU.PC += 2;
    if (read)
       OpAddress = S9xGetWord(OpAddress) | ((OpenBus = S9xGetByte(OpAddress + 2)) << 16);
@@ -83,12 +83,12 @@ static INLINE void AbsoluteIndirect(bool read)
 #ifdef FAST_LSB_WORD_ACCESS
    OpAddress = *(uint16_t*) CPU.PC;
 #else
-   OpAddress = *CPU.PC + (*(CPU.PC + 1) << 8);
+   OpAddress = CPU.PC[0] + (CPU.PC[1] << 8);
 #endif
 #ifndef SA1_OPCODES
    CPU.Cycles += CPU.MemSpeedx2;
 #endif
-   OpenBus = *(CPU.PC + 1);
+   OpenBus = CPU.PC[1];
    CPU.PC += 2;
    OpAddress = S9xGetWord(OpAddress);
    if (read)
@@ -101,10 +101,10 @@ static INLINE void Absolute(bool read)
 #ifdef FAST_LSB_WORD_ACCESS
    OpAddress = *(uint16_t*) CPU.PC + ICPU.ShiftedDB;
 #else
-   OpAddress = *CPU.PC + (*(CPU.PC + 1) << 8) + ICPU.ShiftedDB;
+   OpAddress = CPU.PC[0] + (CPU.PC[1] << 8) + ICPU.ShiftedDB;
 #endif
    if (read)
-      OpenBus = *(CPU.PC + 1);
+      OpenBus = CPU.PC[1];
    CPU.PC += 2;
 #ifndef SA1_OPCODES
    CPU.Cycles += CPU.MemSpeedx2;
@@ -117,14 +117,14 @@ static INLINE void AbsoluteLong(bool read)
    OpAddress = (*(uint32_t*) CPU.PC) & 0xffffff;
 #elif defined FAST_ALIGNED_LSB_WORD_ACCESS
    if (((int32_t) CPU.PC & 1) == 0)
-      OpAddress = (*(uint16_t*) CPU.PC) + (*(CPU.PC + 2) << 16);
+      OpAddress = (*(uint16_t*) CPU.PC) + (CPU.PC[2] << 16);
    else
-      OpAddress = *CPU.PC + ((*(uint16_t*)(CPU.PC + 1)) << 8);
+      OpAddress = *CPU.PC + ((*(uint16_t*) (CPU.PC + 1)) << 8);
 #else
-   OpAddress = *CPU.PC + (*(CPU.PC + 1) << 8) + (*(CPU.PC + 2) << 16);
+   OpAddress = CPU.PC[0] + (CPU.PC[1] << 8) + (CPU.PC[2] << 16);
 #endif
    if (read)
-      OpenBus = *(CPU.PC + 2);
+      OpenBus = CPU.PC[2];
    CPU.PC += 3;
 #ifndef SA1_OPCODES
    CPU.Cycles += CPU.MemSpeedx2 + CPU.MemSpeed;
@@ -210,10 +210,10 @@ static INLINE void AbsoluteIndexedX(bool read)
 #ifdef FAST_LSB_WORD_ACCESS
    OpAddress = ICPU.ShiftedDB + *(uint16_t*) CPU.PC + ICPU.Registers.X.W;
 #else
-   OpAddress = ICPU.ShiftedDB + *CPU.PC + (*(CPU.PC + 1) << 8) + ICPU.Registers.X.W;
+   OpAddress = ICPU.ShiftedDB + CPU.PC[0] + (CPU.PC[1] << 8) + ICPU.Registers.X.W;
 #endif
    if (read)
-      OpenBus = *(CPU.PC + 1);
+      OpenBus = CPU.PC[1];
    CPU.PC += 2;
 #ifndef SA1_OPCODES
    CPU.Cycles += CPU.MemSpeedx2;
@@ -225,10 +225,10 @@ static INLINE void AbsoluteIndexedY(bool read)
 #ifdef FAST_LSB_WORD_ACCESS
    OpAddress = ICPU.ShiftedDB + *(uint16_t*) CPU.PC + ICPU.Registers.Y.W;
 #else
-   OpAddress = ICPU.ShiftedDB + *CPU.PC + (*(CPU.PC + 1) << 8) + ICPU.Registers.Y.W;
+   OpAddress = ICPU.ShiftedDB + CPU.PC[0] + (CPU.PC[1] << 8) + ICPU.Registers.Y.W;
 #endif
    if (read)
-      OpenBus = *(CPU.PC + 1);
+      OpenBus = CPU.PC[1];
    CPU.PC += 2;
 #ifndef SA1_OPCODES
    CPU.Cycles += CPU.MemSpeedx2;
@@ -241,14 +241,14 @@ static INLINE void AbsoluteLongIndexedX(bool read)
     OpAddress = (*(uint32_t*) CPU.PC + ICPU.Registers.X.W) & 0xffffff;
 #elif defined FAST_ALIGNED_LSB_WORD_ACCESS
    if (((int32_t) CPU.PC & 1) == 0)
-       OpAddress = ((*(uint16_t*) CPU.PC) + (*(CPU.PC + 2) << 16) + ICPU.Registers.X.W) & 0xFFFFFF;
+       OpAddress = ((*(uint16_t*) CPU.PC) + (CPU.PC[2] << 16) + ICPU.Registers.X.W) & 0xFFFFFF;
    else
-       OpAddress = (*CPU.PC + ((*(uint16_t*)(CPU.PC + 1)) << 8) + ICPU.Registers.X.W) & 0xFFFFFF;
+       OpAddress = (*CPU.PC + ((*(uint16_t*) (CPU.PC + 1)) << 8) + ICPU.Registers.X.W) & 0xFFFFFF;
 #else
-    OpAddress = (*CPU.PC + (*(CPU.PC + 1) << 8) + (*(CPU.PC + 2) << 16) + ICPU.Registers.X.W) & 0xffffff;
+    OpAddress = (CPU.PC[0] + (CPU.PC[1] << 8) + (CPU.PC[2] << 16) + ICPU.Registers.X.W) & 0xffffff;
 #endif
     if (read)
-       OpenBus = *(CPU.PC + 2);
+       OpenBus = CPU.PC[2];
    CPU.PC += 3;
 #ifndef SA1_OPCODES
     CPU.Cycles += CPU.MemSpeedx2 + CPU.MemSpeed;
diff --git a/source/dma.c b/source/dma.c
index 7effb25..b9f977f 100644
--- a/source/dma.c
+++ b/source/dma.c
@@ -130,14 +130,14 @@ void S9xDoDMA(uint8_t Channel)
                      for (b = 0; b < 2; b++)
                      {
                         uint8_t r = *(q + b);
-                        *(p + 0) = (*(p + 0) << 1) | ((r >> 0) & 1);
-                        *(p + 1) = (*(p + 1) << 1) | ((r >> 1) & 1);
-                        *(p + 0) = (*(p + 0) << 1) | ((r >> 2) & 1);
-                        *(p + 1) = (*(p + 1) << 1) | ((r >> 3) & 1);
-                        *(p + 0) = (*(p + 0) << 1) | ((r >> 4) & 1);
-                        *(p + 1) = (*(p + 1) << 1) | ((r >> 5) & 1);
-                        *(p + 0) = (*(p + 0) << 1) | ((r >> 6) & 1);
-                        *(p + 1) = (*(p + 1) << 1) | ((r >> 7) & 1);
+                        p[0] = (p[0] << 1) | ((r >> 0) & 1);
+                        p[1] = (p[1] << 1) | ((r >> 1) & 1);
+                        p[0] = (p[0] << 1) | ((r >> 2) & 1);
+                        p[1] = (p[1] << 1) | ((r >> 3) & 1);
+                        p[0] = (p[0] << 1) | ((r >> 4) & 1);
+                        p[1] = (p[1] << 1) | ((r >> 5) & 1);
+                        p[0] = (p[0] << 1) | ((r >> 6) & 1);
+                        p[1] = (p[1] << 1) | ((r >> 7) & 1);
                      }
                      p += 2;
                   }
@@ -158,14 +158,14 @@ void S9xDoDMA(uint8_t Channel)
                      for (b = 0; b < 4; b++)
                      {
                         uint8_t r = *(q + b);
-                        *(p +  0) = (*(p +  0) << 1) | ((r >> 0) & 1);
-                        *(p +  1) = (*(p +  1) << 1) | ((r >> 1) & 1);
-                        *(p + 16) = (*(p + 16) << 1) | ((r >> 2) & 1);
-                        *(p + 17) = (*(p + 17) << 1) | ((r >> 3) & 1);
-                        *(p +  0) = (*(p +  0) << 1) | ((r >> 4) & 1);
-                        *(p +  1) = (*(p +  1) << 1) | ((r >> 5) & 1);
-                        *(p + 16) = (*(p + 16) << 1) | ((r >> 6) & 1);
-                        *(p + 17) = (*(p + 17) << 1) | ((r >> 7) & 1);
+                        p[0]  = (p[0]  << 1) | ((r >> 0) & 1);
+                        p[1]  = (p[1]  << 1) | ((r >> 1) & 1);
+                        p[16] = (p[16] << 1) | ((r >> 2) & 1);
+                        p[17] = (p[17] << 1) | ((r >> 3) & 1);
+                        p[0]  = (p[0]  << 1) | ((r >> 4) & 1);
+                        p[1]  = (p[1]  << 1) | ((r >> 5) & 1);
+                        p[16] = (p[16] << 1) | ((r >> 6) & 1);
+                        p[17] = (p[17] << 1) | ((r >> 7) & 1);
                      }
                      p += 2;
                   }
@@ -187,14 +187,14 @@ void S9xDoDMA(uint8_t Channel)
                      for (b = 0; b < 8; b++)
                      {
                         uint8_t r = *(q + b);
-                        *(p +  0) = (*(p +  0) << 1) | ((r >> 0) & 1);
-                        *(p +  1) = (*(p +  1) << 1) | ((r >> 1) & 1);
-                        *(p + 16) = (*(p + 16) << 1) | ((r >> 2) & 1);
-                        *(p + 17) = (*(p + 17) << 1) | ((r >> 3) & 1);
-                        *(p + 32) = (*(p + 32) << 1) | ((r >> 4) & 1);
-                        *(p + 33) = (*(p + 33) << 1) | ((r >> 5) & 1);
-                        *(p + 48) = (*(p + 48) << 1) | ((r >> 6) & 1);
-                        *(p + 49) = (*(p + 49) << 1) | ((r >> 7) & 1);
+                        p[0]  = (p[0]  << 1) | ((r >> 0) & 1);
+                        p[1]  = (p[1]  << 1) | ((r >> 1) & 1);
+                        p[16] = (p[16] << 1) | ((r >> 2) & 1);
+                        p[17] = (p[17] << 1) | ((r >> 3) & 1);
+                        p[32] = (p[32] << 1) | ((r >> 4) & 1);
+                        p[33] = (p[33] << 1) | ((r >> 5) & 1);
+                        p[48] = (p[48] << 1) | ((r >> 6) & 1);
+                        p[49] = (p[49] << 1) | ((r >> 7) & 1);
                      }
                      p += 2;
                   }
diff --git a/source/fxemu.c b/source/fxemu.c
index dd798fd..9d62b05 100644
--- a/source/fxemu.c
+++ b/source/fxemu.c
@@ -6,6 +6,8 @@
 #include <string.h>
 #include <stdio.h>
 
+#include <retro_inline.h>
+
 /* The FxChip Emulator's internal variables */
 FxRegs_s GSU; /* This will be initialized when loading a ROM */
 
@@ -27,13 +29,10 @@ void fx_updateRamBank(uint8_t Byte)
    GSU.pvRamBank = GSU.apvRamBank[Byte & 0x3];
 }
 
-static void fx_readRegisterSpaceForCheck(void)
+static INLINE void fx_readRegisterSpaceForCheck(void)
 {
    R15 = GSU.pvRegisters[30];
    R15 |= ((uint32_t) GSU.pvRegisters[31]) << 8;
-   GSU.vStatusReg = (uint32_t) GSU.pvRegisters[GSU_SFR];
-   GSU.vStatusReg |= ((uint32_t) GSU.pvRegisters[GSU_SFR + 1]) << 8;
-   GSU.vPrgBankReg = (uint32_t) GSU.pvRegisters[GSU_PBR];
 }
 
 static void fx_readRegisterSpaceForUse(void)
@@ -52,6 +51,9 @@ static void fx_readRegisterSpaceForUse(void)
 
    /* Update other registers */
    p = GSU.pvRegisters;
+   GSU.vStatusReg = (uint32_t) GSU.pvRegisters[GSU_SFR];
+   GSU.vStatusReg |= ((uint32_t) GSU.pvRegisters[GSU_SFR + 1]) << 8;
+   GSU.vPrgBankReg = (uint32_t) GSU.pvRegisters[GSU_PBR];
    GSU.vRomBankReg = (uint32_t)p[GSU_ROMBR];
    GSU.vRamBankReg = ((uint32_t)p[GSU_RAMBR]) & (FX_RAM_BANKS - 1);
    GSU.vCacheBaseReg = (uint32_t)p[GSU_CBR];
@@ -75,9 +77,9 @@ static void fx_readRegisterSpaceForUse(void)
    GSU.vScreenHeight = GSU.vScreenRealHeight = avHeight[i];
    GSU.vMode = p[GSU_SCMR] & 0x03;
    if (i == 3)
-      GSU.vScreenSize = (256 / 8) * (256 / 8) * 32;
+      GSU.vScreenSize = 32768;
    else
-      GSU.vScreenSize = (GSU.vScreenHeight / 8) * (256 / 8) * avMult[GSU.vMode];
+      GSU.vScreenSize = GSU.vScreenHeight * 4 * avMult[GSU.vMode];
    if (GSU.vPlotOptionReg & 0x10)
       GSU.vScreenHeight = 256; /* OBJ Mode (for drawing into sprites) */
    if (GSU.pvScreenBase + GSU.vScreenSize > GSU.pvRam + (GSU.nRamBanks * 65536))
@@ -90,7 +92,8 @@ static void fx_readRegisterSpaceForUse(void)
    fx_apfOpcodeTable[0x24c] = GSU.pfPlot;
    fx_apfOpcodeTable[0x34c] = GSU.pfRpix;
 
-   fx_computeScreenPointers();
+   if(GSU.vMode != GSU.vPrevMode || GSU.vPrevScreenHeight != GSU.vScreenHeight || GSU.vSCBRDirty)
+      fx_computeScreenPointers();
 }
 
 void fx_dirtySCBR(void)
@@ -100,135 +103,52 @@ void fx_dirtySCBR(void)
 
 void fx_computeScreenPointers(void)
 {
-   if (GSU.vMode != GSU.vPrevMode || GSU.vPrevScreenHeight != GSU.vScreenHeight || GSU.vSCBRDirty)
+   int32_t i, j, condition, mask, result;
+   uint32_t apvIncrement, vMode, xIncrement;
+   GSU.vSCBRDirty = false;
+
+   /* Make a list of pointers to the start of each screen column*/
+   vMode = GSU.vMode;
+   condition = vMode - 2;
+   mask = (condition | -condition) >> 31;
+   result = (vMode & mask) | (3 & ~mask);
+   vMode = result + 1;
+   GSU.x[0] = 0;
+   GSU.apvScreen[0] = GSU.pvScreenBase;
+   apvIncrement = vMode << 4;
+
+   if(GSU.vScreenHeight == 256)
    {
-      int32_t i;
-      GSU.vSCBRDirty = false;
+      GSU.x[16] = vMode << 12;
+      GSU.apvScreen[16] = GSU.pvScreenBase + (vMode << 13);
+      apvIncrement <<= 4;
+      xIncrement = vMode << 4;
 
-      /* Make a list of pointers to the start of each screen column */
-      switch (GSU.vScreenHeight)
+      for(i = 1, j = 17 ; i < 16 ; i++, j++)
+      {
+         GSU.x[i] = GSU.x[i - 1] + xIncrement;
+         GSU.apvScreen[i] = GSU.apvScreen[i - 1] + apvIncrement;
+         GSU.x[j] = GSU.x[j - 1] + xIncrement;
+         GSU.apvScreen[j] = GSU.apvScreen[j - 1] + apvIncrement;
+      }
+   }
+   else
+   {
+      xIncrement = (vMode * GSU.vScreenHeight) << 1;
+      for(i = 1 ; i < 32 ; i++)
       {
-      case 128:
-         switch (GSU.vMode)
-         {
-         case 0:
-            for (i = 0; i < 32; i++)
-            {
-               GSU.apvScreen[i] = GSU.pvScreenBase + (i << 4);
-               GSU.x[i] = i << 8;
-            }
-            break;
-         case 1:
-            for (i = 0; i < 32; i++)
-            {
-               GSU.apvScreen[i] = GSU.pvScreenBase + (i << 5);
-               GSU.x[i] = i << 9;
-            }
-            break;
-         case 2:
-         case 3:
-            for (i = 0; i < 32; i++)
-            {
-               GSU.apvScreen[i] = GSU.pvScreenBase + (i << 6);
-               GSU.x[i] = i << 10;
-            }
-            break;
-         }
-         break;
-      case 160:
-         switch (GSU.vMode)
-         {
-         case 0:
-            for (i = 0; i < 32; i++)
-            {
-               GSU.apvScreen[i] = GSU.pvScreenBase + (i << 4);
-               GSU.x[i] = (i << 8) + (i << 6);
-            }
-            break;
-         case 1:
-            for (i = 0; i < 32; i++)
-            {
-               GSU.apvScreen[i] = GSU.pvScreenBase + (i << 5);
-               GSU.x[i] = (i << 9) + (i << 7);
-            }
-            break;
-         case 2:
-         case 3:
-            for (i = 0; i < 32; i++)
-            {
-               GSU.apvScreen[i] = GSU.pvScreenBase + (i << 6);
-               GSU.x[i] = (i << 10) + (i << 8);
-            }
-            break;
-         }
-         break;
-      case 192:
-         switch (GSU.vMode)
-         {
-         case 0:
-            for (i = 0; i < 32; i++)
-            {
-               GSU.apvScreen[i] = GSU.pvScreenBase + (i << 4);
-               GSU.x[i] = (i << 8) + (i << 7);
-            }
-            break;
-         case 1:
-            for (i = 0; i < 32; i++)
-            {
-               GSU.apvScreen[i] = GSU.pvScreenBase + (i << 5);
-               GSU.x[i] = (i << 9) + (i << 8);
-            }
-            break;
-         case 2:
-         case 3:
-            for (i = 0; i < 32; i++)
-            {
-               GSU.apvScreen[i] = GSU.pvScreenBase + (i << 6);
-               GSU.x[i] = (i << 10) + (i << 9);
-            }
-            break;
-         }
-         break;
-      case 256:
-         switch (GSU.vMode)
-         {
-         case 0:
-            for (i = 0; i < 32; i++)
-            {
-               GSU.apvScreen[i] = GSU.pvScreenBase + ((i & 0x10) << 9) + ((i & 0xf) << 8);
-               GSU.x[i] = ((i & 0x10) << 8) + ((i & 0xf) << 4);
-            }
-            break;
-         case 1:
-            for (i = 0; i < 32; i++)
-            {
-               GSU.apvScreen[i] = GSU.pvScreenBase + ((i & 0x10) << 10) + ((i & 0xf) << 9);
-               GSU.x[i] = ((i & 0x10) << 9) + ((i & 0xf) << 5);
-            }
-            break;
-         case 2:
-         case 3:
-            for (i = 0; i < 32; i++)
-            {
-               GSU.apvScreen[i] = GSU.pvScreenBase + ((i & 0x10) << 11) + ((i & 0xf) << 10);
-               GSU.x[i] = ((i & 0x10) << 10) + ((i & 0xf) << 6);
-            }
-            break;
-         }
-         break;
+         GSU.x[i] = GSU.x[i - 1] + xIncrement;
+         GSU.apvScreen[i] = GSU.apvScreen[i - 1] + apvIncrement;
       }
-      GSU.vPrevMode = GSU.vMode;
-      GSU.vPrevScreenHeight = GSU.vScreenHeight;
    }
+   GSU.vPrevMode = GSU.vMode;
+   GSU.vPrevScreenHeight = GSU.vScreenHeight;
 }
 
-static void fx_writeRegisterSpaceAfterCheck(void)
+static INLINE void fx_writeRegisterSpaceAfterCheck(void)
 {
    GSU.pvRegisters[30] = (uint8_t) R15;
    GSU.pvRegisters[31] = (uint8_t) (R15 >> 8);
-   GSU.pvRegisters[GSU_SFR] = (uint8_t) GSU.vStatusReg;
-   GSU.pvRegisters[GSU_SFR + 1] = (uint8_t) (GSU.vStatusReg >> 8);
-   GSU.pvRegisters[GSU_PBR] = (uint8_t) GSU.vPrgBankReg;
 }
 
 static void fx_writeRegisterSpaceAfterUse(void)
@@ -260,6 +180,9 @@ static void fx_writeRegisterSpaceAfterUse(void)
       CF(CY);
 
    p = GSU.pvRegisters;
+   p[GSU_SFR] = (uint8_t) GSU.vStatusReg;
+   p[GSU_SFR + 1] = (uint8_t) (GSU.vStatusReg >> 8);
+   p[GSU_PBR] = (uint8_t) GSU.vPrgBankReg;
    p[GSU_ROMBR] = (uint8_t)GSU.vRomBankReg;
    p[GSU_RAMBR] = (uint8_t)GSU.vRamBankReg;
    p[GSU_CBR] = (uint8_t)GSU.vCacheBaseReg;
@@ -342,7 +265,7 @@ static bool fx_checkStartAddress(void)
       return false;
 
    /* Check if we're in RAM and the RAN flag is not set */
-   if (GSU.vPrgBankReg >= 0x70 && GSU.vPrgBankReg <= 0x73 && !(SCMR & (1 << 3)))
+   if (GSU.vPrgBankReg >= 0x70 && !(SCMR & (1 << 3)))
       return false;
 
    /* If not, we're in ROM, so check if the RON flag is set */
diff --git a/source/gfx.c b/source/gfx.c
index f3b59a5..6dd9ff5 100644
--- a/source/gfx.c
+++ b/source/gfx.c
@@ -673,10 +673,9 @@ void S9xSetupOBJ(void)
       for (Y = 1; Y < SNES_HEIGHT_EXTENDED; Y++)
          GFX.OBJLines[Y].RTOFlags |= GFX.OBJLines[Y - 1].RTOFlags;
    }
-   else
+   else /* evil FirstSprite+Y case */
    {
       int32_t j, Y;
-      /* evil FirstSprite+Y case */
       /* First, find out which sprites are on which lines */
       uint8_t OBJOnLine[SNES_HEIGHT_EXTENDED][128];
       /* We only initialise this per line, as needed. [Neb]
@@ -2080,7 +2079,7 @@ static void DrawBackground(uint32_t BGMode, uint32_t bg, uint8_t Z1, uint8_t Z2)
           int32_t X = ((AA + BB) >> 8) & 0x3ff; \
           int32_t Y = ((CC + DD) >> 8) & 0x3ff; \
           uint8_t *TileData = VRAM1 + (Memory.VRAM[((Y & ~7) << 5) + ((X >> 2) & ~1)] << 7); \
-          uint32_t b = *(TileData + ((Y & 7) << 4) + ((X & 7) << 1)); \
+          uint32_t b = TileData[((Y & 7) << 4) + ((X & 7) << 1)]; \
           GFX.Z1 = Mode7Depths [(b & GFX.Mode7PriorityMask) >> 7]; \
           if (GFX.Z1 > *d && (b & GFX.Mode7Mask) ) \
           { \
@@ -2100,7 +2099,7 @@ static void DrawBackground(uint32_t BGMode, uint32_t bg, uint8_t Z1, uint8_t Z2)
           if (((X | Y) & ~0x3ff) == 0) \
           { \
          uint8_t *TileData = VRAM1 + (Memory.VRAM[((Y & ~7) << 5) + ((X >> 2) & ~1)] << 7); \
-         uint32_t b = *(TileData + ((Y & 7) << 4) + ((X & 7) << 1)); \
+         uint32_t b = TileData[((Y & 7) << 4) + ((X & 7) << 1)]; \
          GFX.Z1 = Mode7Depths [(b & GFX.Mode7PriorityMask) >> 7]; \
          if (GFX.Z1 > *d && (b & GFX.Mode7Mask) ) \
          { \
@@ -2115,7 +2114,7 @@ static void DrawBackground(uint32_t BGMode, uint32_t bg, uint8_t Z1, uint8_t Z2)
              uint32_t b; \
              X = (x + HOffset) & 7; \
              Y = (yy + CentreY) & 7; \
-             b = *(VRAM1 + ((Y & 7) << 4) + ((X & 7) << 1)); \
+             b = VRAM1[((Y & 7) << 4) + ((X & 7) << 1)]; \
              GFX.Z1 = Mode7Depths [(b & GFX.Mode7PriorityMask) >> 7]; \
              if (GFX.Z1 > *d && (b & GFX.Mode7Mask) ) \
              { \
@@ -2284,7 +2283,7 @@ static void DrawBGMode7Background16Sub1_2(uint8_t * Screen, int32_t bg)
                         int32_t X = ((AA + BB) >> 8) & 0x3ff; \
                         int32_t Y = (DD >> 8) & 0x3ff; \
                         uint8_t *TileData = VRAM1 + (Memory.VRAM[((Y & ~7) << 5) + ((X >> 2) & ~1)] << 7); \
-                        b = *(TileData + ((Y & 7) << 4) + ((X & 7) << 1)); \
+                        b = TileData[((Y & 7) << 4) + ((X & 7) << 1)]; \
          GFX.Z1 = Mode7Depths [(b & GFX.Mode7PriorityMask) >> 7]; \
                         if (GFX.Z1 > *d && (b & GFX.Mode7Mask) ) \
                         { \
@@ -2306,7 +2305,7 @@ static void DrawBGMode7Background16Sub1_2(uint8_t * Screen, int32_t bg)
                         if (((X | Y) & ~0x3ff) == 0) \
                         { \
                             uint8_t *TileData = VRAM1 + (Memory.VRAM[((Y & ~7) << 5) + ((X >> 2) & ~1)] << 7); \
-             b = *(TileData + ((Y & 7) << 4) + ((X & 7) << 1)); \
+             b = TileData[((Y & 7) << 4) + ((X & 7) << 1)]; \
              GFX.Z1 = Mode7Depths [(b & GFX.Mode7PriorityMask) >> 7]; \
                             if (GFX.Z1 > *d && (b & GFX.Mode7Mask) ) \
                             { \
@@ -2322,7 +2321,7 @@ static void DrawBGMode7Background16Sub1_2(uint8_t * Screen, int32_t bg)
                             X = (x + HOffset) & 7; \
                             Y = (yy + CentreY) & 7; \
              TileData = VRAM1 + (Memory.VRAM[((Y & ~7) << 5) + ((X >> 2) & ~1)] << 7); \
-             b = *(TileData + ((Y & 7) << 4) + ((X & 7) << 1)); \
+             b = TileData[((Y & 7) << 4) + ((X & 7) << 1)]; \
              GFX.Z1 = Mode7Depths [(b & GFX.Mode7PriorityMask) >> 7]; \
                             if (GFX.Z1 > *d && (b & GFX.Mode7Mask) ) \
                             { \
@@ -2360,7 +2359,7 @@ static void DrawBGMode7Background16Sub1_2(uint8_t * Screen, int32_t bg)
                         uint32_t X = xPix & 0x3ff; \
                         uint32_t Y = yPix & 0x3ff; \
                         uint8_t *TileData = VRAM1 + (Memory.VRAM[((Y & ~7) << 5) + ((X >> 2) & ~1)] << 7); \
-         b = *(TileData + ((Y & 7) << 4) + ((X & 7) << 1)); \
+         b = TileData[((Y & 7) << 4) + ((X & 7) << 1)]; \
          GFX.Z1 = Mode7Depths [(b & GFX.Mode7PriorityMask) >> 7]; \
                         if (GFX.Z1 > *d && (b & GFX.Mode7Mask) ) \
                         { \
@@ -2375,13 +2374,13 @@ static void DrawBGMode7Background16Sub1_2(uint8_t * Screen, int32_t bg)
                             uint8_t *TileData01 = VRAM1 + (Memory.VRAM[((Y01 & ~7) << 5) + ((X >> 2) & ~1)] << 7); \
                             p1 = COLORFUNC; \
                             p1 = (p1 & FIRST_THIRD_COLOR_MASK) | ((p1 & SECOND_COLOR_MASK) << 16); \
-                            b = *(TileData10 + ((Y & 7) << 4) + ((X10 & 7) << 1)); \
+                            b = TileData10[((Y & 7) << 4) + ((X10 & 7) << 1)]; \
                             p2 = COLORFUNC; \
                             p2 = (p2 & FIRST_THIRD_COLOR_MASK) | ((p2 & SECOND_COLOR_MASK) << 16); \
-                            b = *(TileData11 + ((Y01 & 7) << 4) + ((X10 & 7) << 1)); \
+                            b = TileData11[((Y01 & 7) << 4) + ((X10 & 7) << 1)]; \
                             p4 = COLORFUNC; \
                             p4 = (p4 & FIRST_THIRD_COLOR_MASK) | ((p4 & SECOND_COLOR_MASK) << 16); \
-                            b = *(TileData01 + ((Y01 & 7) << 4) + ((X & 7) << 1)); \
+                            b = TileData01[((Y01 & 7) << 4) + ((X & 7) << 1)]; \
                             p3 = COLORFUNC; \
                             p3 = (p3 & FIRST_THIRD_COLOR_MASK) | ((p3 & SECOND_COLOR_MASK) << 16); \
                             /* Xdel, Ydel: position (in 1/32nds) between the points */ \
@@ -2439,7 +2438,7 @@ static void DrawBGMode7Background16Sub1_2(uint8_t * Screen, int32_t bg)
                         uint32_t X = ((AA + BB) >> 8) & 0x3ff; \
                         uint32_t Y = ((CC + DD) >> 8) & 0x3ff; \
                         uint8_t *TileData = VRAM1 + (Memory.VRAM[((Y & ~7) << 5) + ((X >> 2) & ~1)] << 7); \
-         b = *(TileData + ((Y & 7) << 4) + ((X & 7) << 1)); \
+         b = TileData[((Y & 7) << 4) + ((X & 7) << 1)]; \
          GFX.Z1 = Mode7Depths [(b & GFX.Mode7PriorityMask) >> 7]; \
                         if (GFX.Z1 > *d && (b & GFX.Mode7Mask) ) \
                         { \
@@ -2456,11 +2455,11 @@ static void DrawBGMode7Background16Sub1_2(uint8_t * Screen, int32_t bg)
                             uint8_t *TileData01 = VRAM1 + (Memory.VRAM[((Y01 & ~7) << 5) + ((X01 >> 2) & ~1)] << 7); \
                             uint8_t *TileData11 = VRAM1 + (Memory.VRAM[((Y11 & ~7) << 5) + ((X11 >> 2) & ~1)] << 7); \
                             p1 = COLORFUNC; \
-                            b = *(TileData10 + ((Y10 & 7) << 4) + ((X10 & 7) << 1)); \
+                            b = TileData10[((Y10 & 7) << 4) + ((X10 & 7) << 1)]; \
                             p2 = COLORFUNC; \
-                            b = *(TileData01 + ((Y01 & 7) << 4) + ((X01 & 7) << 1)); \
+                            b = TileData01[((Y01 & 7) << 4) + ((X01 & 7) << 1)]; \
                             p3 = COLORFUNC; \
-                            b = *(TileData11 + ((Y11 & 7) << 4) + ((X11 & 7) << 1)); \
+                            b = TileData11[((Y11 & 7) << 4) + ((X11 & 7) << 1)]; \
                             p4 = COLORFUNC; \
                             theColor = Q_INTERPOLATE(p1, p2, p3, p4); \
                             *p = (FUNC) | ALPHA_BITS_MASK; \
@@ -2485,7 +2484,7 @@ static void DrawBGMode7Background16Sub1_2(uint8_t * Screen, int32_t bg)
                     if (((X | Y) & ~0x3ff) == 0) \
                     { \
                         uint8_t *TileData = VRAM1 + (Memory.VRAM[((Y & ~7) << 5) + ((X >> 2) & ~1)] << 7); \
-         b = *(TileData + ((Y & 7) << 4) + ((X & 7) << 1)); \
+         b = TileData[((Y & 7) << 4) + ((X & 7) << 1)]; \
          GFX.Z1 = Mode7Depths [(b & GFX.Mode7PriorityMask) >> 7]; \
                         if (GFX.Z1 > *d && (b & GFX.Mode7Mask) ) \
                         { \
@@ -2501,13 +2500,13 @@ static void DrawBGMode7Background16Sub1_2(uint8_t * Screen, int32_t bg)
                             uint8_t *TileData01 = VRAM1 + (Memory.VRAM[((Y01 & ~7) << 5) + ((X >> 2) & ~1)] << 7); \
                             p1 = COLORFUNC; \
                             p1 = (p1 & FIRST_THIRD_COLOR_MASK) | ((p1 & SECOND_COLOR_MASK) << 16); \
-                            b = *(TileData10 + ((Y & 7) << 4) + ((X10 & 7) << 1)); \
+                            b = TileData10[((Y & 7) << 4) + ((X10 & 7) << 1)]; \
                             p2 = COLORFUNC; \
                             p2 = (p2 & FIRST_THIRD_COLOR_MASK) | ((p2 & SECOND_COLOR_MASK) << 16); \
-                            b = *(TileData11 + ((Y01 & 7) << 4) + ((X10 & 7) << 1)); \
+                            b = TileData11[((Y01 & 7) << 4) + ((X10 & 7) << 1)]; \
                             p4 = COLORFUNC; \
                             p4 = (p4 & FIRST_THIRD_COLOR_MASK) | ((p4 & SECOND_COLOR_MASK) << 16); \
-                            b = *(TileData01 + ((Y01 & 7) << 4) + ((X & 7) << 1)); \
+                            b = TileData01[((Y01 & 7) << 4) + ((X & 7) << 1)]; \
                             p3 = COLORFUNC; \
                             p3 = (p3 & FIRST_THIRD_COLOR_MASK) | ((p3 & SECOND_COLOR_MASK) << 16); \
                             /* Xdel, Ydel: position (in 1/32nds) between the points */ \
@@ -2533,7 +2532,7 @@ static void DrawBGMode7Background16Sub1_2(uint8_t * Screen, int32_t bg)
                         { \
                             X = (x + HOffset) & 7; \
                             Y = (yy + CentreY) & 7; \
-             b = *(VRAM1 + ((Y & 7) << 4) + ((X & 7) << 1)); \
+             b = VRAM1[((Y & 7) << 4) + ((X & 7) << 1)]; \
              GFX.Z1 = Mode7Depths [(b & GFX.Mode7PriorityMask) >> 7]; \
                             if (GFX.Z1 > *d && (b & GFX.Mode7Mask) ) \
                             { \
@@ -2768,10 +2767,10 @@ void S9xUpdateScreen(void)
          for (y = 0; y < starty; y++)
          {
             int32_t x;
-            uint16_t* p = (uint16_t*)(GFX.Screen + y * GFX.Pitch2) + 255;
-            uint16_t* q = (uint16_t*)(GFX.Screen + y * GFX.Pitch2) + 510;
+            uint16_t* p = (uint16_t*) (GFX.Screen + y * GFX.Pitch2) + 255;
+            uint16_t* q = (uint16_t*) p + 255;
             for (x = 255; x >= 0; x--, p--, q -= 2)
-               * q = *(q + 1) = *p;
+               q[0] = q[1] = p[0];
          }
          IPPU.DoubleWidthPixels = true;
          IPPU.HalfWidthPixels = false;
@@ -3219,9 +3218,9 @@ void S9xUpdateScreen(void)
       {
          int32_t x;
          uint16_t* p = (uint16_t*)(GFX.Screen + y * GFX.Pitch2) + 255;
-         uint16_t* q = (uint16_t*)(GFX.Screen + y * GFX.Pitch2) + 510;
+         uint16_t* q = p + 255;
          for (x = 255; x >= 0; x--, p--, q -= 2)
-            * q = *(q + 1) = *p;
+            q[0] = q[1] = p[0];
       }
    }
 
diff --git a/source/sa1.c b/source/sa1.c
index b17b562..afa3021 100644
--- a/source/sa1.c
+++ b/source/sa1.c
@@ -151,7 +151,7 @@ void S9xSA1SetByte(uint8_t byte, uint32_t address)
 
    if (SetAddress >= (uint8_t*) MAP_LAST)
    {
-      *(SetAddress + (address & 0xffff)) = byte;
+      SetAddress[address & 0xffff] = byte;
       return;
    }
 
@@ -162,10 +162,10 @@ void S9xSA1SetByte(uint8_t byte, uint32_t address)
       return;
    case MAP_SA1RAM:
    case MAP_LOROM_SRAM:
-      *(Memory.SRAM + (address & 0xffff)) = byte;
+      Memory.SRAM[address & 0xffff] = byte;
       return;
    case MAP_BWRAM:
-      *(SA1.BWRAM + ((address & 0x7fff) - 0x6000)) = byte;
+      SA1.BWRAM[(address & 0x7fff) - 0x6000] = byte;
       return;
    case MAP_BWRAM_BITMAP:
       address -= 0x600000;
@@ -186,7 +186,7 @@ void S9xSA1SetByte(uint8_t byte, uint32_t address)
       address = (address & 0xffff) - 0x6000;
       if (SA1.VirtualBitmapFormat == 2)
       {
-         uint8_t* ptr = &SA1.BWRAM [(address >> 2) & 0xffff];
+         uint8_t* ptr = &SA1.BWRAM[(address >> 2) & 0xffff];
          *ptr &= ~(3 << ((address & 3) << 1));
          *ptr |= (byte & 3) << ((address & 3) << 1);
       }
@@ -264,8 +264,8 @@ void S9xSetSA1MemMap(uint32_t which1, uint8_t map)
 
    for (c = 0; c < 0x200; c += 16)
    {
-      /*Code from Snes9x 1.54.1 - This allows Super Mario World VLDC 9 hack to load.
-        Conversion to int is needed here - map is promoted but which1 is not */
+      /* Code from Snes9x 1.54.1 - This allows Super Mario World VLDC 9 hack to load.
+         Conversion to int is needed here - map is promoted but which1 is not */
       int32_t offset = (((map & 0x80) ? map : which1) & 7) * 0x100000 + (c << 11) - 0x8000;
       block = &Memory.ROM [offset];
       for (i = c + 8; i < c + 16; i++)
@@ -532,35 +532,60 @@ void S9xSetSA1(uint8_t byte, uint32_t address)
    Memory.FillRAM[address] = byte;
 }
 
-static void S9xSA1CharConv2()
+static void S9xSA1CharConv2(void)
 {
-   uint32_t dest = Memory.FillRAM [0x2235] | (Memory.FillRAM [0x2236] << 8);
+   uint32_t dest = Memory.FillRAM[0x2235] | (Memory.FillRAM[0x2236] << 8);
    uint32_t offset = (SA1.in_char_dma & 7) ? 0 : 1;
-   int32_t depth = (Memory.FillRAM [0x2231] & 3) == 0 ? 8 : (Memory.FillRAM [0x2231] & 3) == 1 ? 4 : 2;
-   int32_t bytes_per_char = 8 * depth;
-   uint8_t* p = &Memory.FillRAM [0x3000] + dest + offset * bytes_per_char;
-   uint8_t* q = &Memory.ROM [MAX_ROM_SIZE - 0x10000] + offset * 64;
+   int32_t depthX8 = (Memory.FillRAM[0x2231] & 3) == 0 ? 64 : (Memory.FillRAM[0x2231] & 3) == 1 ? 32 : 16;
+   uint8_t* p = &Memory.FillRAM[0x3000] + (dest & 0x7ff) + offset * depthX8;
+   uint8_t* q = &Memory.ROM[MAX_ROM_SIZE - 0x10000] + offset * 64;
 
-   if (depth == 8)
+   switch(depthX8)
    {
-      int32_t l;
-      for (l = 0; l < 8; l++, q += 8)
-      {
-         int32_t b;
-         for (b = 0; b < 8; b++)
+      case 16:
+         for (int l = 0; l < 8; l++, q += 8)
          {
-            uint8_t r = *(q + b);
-            *(p +  0) = (*(p +  0) << 1) | ((r >> 0) & 1);
-            *(p +  1) = (*(p +  1) << 1) | ((r >> 1) & 1);
-            *(p + 16) = (*(p + 16) << 1) | ((r >> 2) & 1);
-            *(p + 17) = (*(p + 17) << 1) | ((r >> 3) & 1);
-            *(p + 32) = (*(p + 32) << 1) | ((r >> 4) & 1);
-            *(p + 33) = (*(p + 33) << 1) | ((r >> 5) & 1);
-            *(p + 48) = (*(p + 48) << 1) | ((r >> 6) & 1);
-            *(p + 49) = (*(p + 49) << 1) | ((r >> 7) & 1);
+            for (int b = 0; b < 8; b++)
+            {
+               uint8_t r = q[b];
+               p[0] = (p[0] << 1) | ((r >> 0) & 1);
+               p[1] = (p[1] << 1) | ((r >> 1) & 1);
+            }
+            p += 2;
          }
-         p += 2;
-      }
+         break;
+      case 32:
+         for (int l = 0; l < 8; l++, q += 8)
+         {
+            for (int b = 0; b < 8; b++)
+            {
+               uint8_t r = q[b];
+               p[0]  = (p[0]  << 1) | ((r >> 0) & 1);
+               p[1]  = (p[1]  << 1) | ((r >> 1) & 1);
+               p[16] = (p[16] << 1) | ((r >> 2) & 1);
+               p[17] = (p[17] << 1) | ((r >> 3) & 1);
+            }
+            p += 2;
+         }
+         break;
+      case 64:
+         for (int l = 0; l < 8; l++, q += 8)
+         {
+            for (int b = 0; b < 8; b++)
+            {
+               uint8_t r = q[b];
+               p[0]  = (p[0]  << 1) | ((r >> 0) & 1);
+               p[1]  = (p[1]  << 1) | ((r >> 1) & 1);
+               p[16] = (p[16] << 1) | ((r >> 2) & 1);
+               p[17] = (p[17] << 1) | ((r >> 3) & 1);
+               p[32] = (p[32] << 1) | ((r >> 4) & 1);
+               p[33] = (p[33] << 1) | ((r >> 5) & 1);
+               p[48] = (p[48] << 1) | ((r >> 6) & 1);
+               p[49] = (p[49] << 1) | ((r >> 7) & 1);
+            }
+            p += 2;
+         }
+         break;
    }
 }
 
diff --git a/source/soundux.c b/source/soundux.c
index 0de7f4b..ebf6915 100644
--- a/source/soundux.c
+++ b/source/soundux.c
@@ -391,8 +391,8 @@ void DecodeBlock(Channel* ch)
          *raw++ = ((int32_t) sample1 << shift);
          *raw++ = ((int32_t) sample2 << shift);
       }
-      prev1 = *(raw - 2);
-      prev0 = *(raw - 1);
+      prev1 = raw[-2];
+      prev0 = raw[-1];
       break;
    case 1:
       for (i = 8; i != 0; i--)
diff --git a/source/spc700.c b/source/spc700.c
index c5ea4ef..f16ba5b 100644
--- a/source/spc700.c
+++ b/source/spc700.c
@@ -111,14 +111,14 @@ uint32_t Work32 = 0;
 
 #define Pop(b) \
     IAPU.Registers.S++; \
-    (b) = *(IAPU.RAM + 0x100 + IAPU.Registers.S)
+    (b) = IAPU.RAM[0x100 + IAPU.Registers.S]
 
 #ifdef FAST_LSB_WORD_ACCESS
 #define PushW(w) \
     if (IAPU.Registers.S == 0) \
     {\
-       *(IAPU.RAM + 0x1ff) = (w); \
-       *(IAPU.RAM + 0x100) = ((w) >> 8); \
+       IAPU.RAM[0x1ff] = (w); \
+       IAPU.RAM[0x100] = ((w) >> 8); \
     } \
     else \
        *(uint16_t *) (IAPU.RAM + 0xff + IAPU.Registers.S) = w; \
@@ -127,21 +127,21 @@ uint32_t Work32 = 0;
 #define PopW(w) \
     IAPU.Registers.S += 2; \
     if (IAPU.Registers.S == 0) \
-       (w) = *(IAPU.RAM + 0x1ff) | (*(IAPU.RAM + 0x100) << 8); \
+       (w) = IAPU.RAM[0x1ff] | (IAPU.RAM[0x100] << 8); \
     else \
        (w) = *(uint16_t *) (IAPU.RAM + 0xff + IAPU.Registers.S)
 #else
 #define PushW(w) \
-    *(IAPU.RAM + 0xff + IAPU.Registers.S) = w; \
-    *(IAPU.RAM + 0x100 + IAPU.Registers.S) = ((w) >> 8); \
+    IAPU.RAM[0xff + IAPU.Registers.S] = w; \
+    IAPU.RAM[0x100 + IAPU.Registers.S] = ((w) >> 8); \
     IAPU.Registers.S -= 2
 
 #define PopW(w) \
     IAPU.Registers.S += 2; \
     if(IAPU.Registers.S == 0) \
-       (w) = *(IAPU.RAM + 0x1ff) | (*(IAPU.RAM + 0x100) << 8); \
+       (w) = IAPU.RAM[0x1ff] | (IAPU.RAM[0x100] << 8); \
     else \
-       (w) = *(IAPU.RAM + 0xff + IAPU.Registers.S) + (*(IAPU.RAM + 0x100 + IAPU.Registers.S) << 8)
+       (w) = IAPU.RAM[0xff + IAPU.Registers.S] + (IAPU.RAM[0x100 + IAPU.Registers.S] << 8)
 #endif
 
 #define Relative() \
@@ -174,8 +174,8 @@ uint32_t Work32 = 0;
     IAPU.Address = *(uint16_t *) (IAPU.DirectPage + OP1) + IAPU.Registers.YA.B.Y;
 #else
 #define IndexedXIndirect() \
-    IAPU.Address = *(IAPU.DirectPage + ((OP1 + IAPU.Registers.X) & 0xff)) + \
-       (*(IAPU.DirectPage + ((OP1 + IAPU.Registers.X + 1) & 0xff)) << 8);
+    IAPU.Address = IAPU.DirectPage[(OP1 + IAPU.Registers.X) & 0xff] + \
+       (IAPU.DirectPage[(OP1 + IAPU.Registers.X + 1) & 0xff] << 8);
 
 #define Absolute() \
     IAPU.Address = OP1 + (OP2 << 8);
@@ -192,8 +192,8 @@ uint32_t Work32 = 0;
     IAPU.Address &= 0x1fff;
 
 #define IndirectIndexedY() \
-    IAPU.Address = *(IAPU.DirectPage + OP1) + \
-       (*(IAPU.DirectPage + OP1 + 1) << 8) + \
+    IAPU.Address = IAPU.DirectPage[OP1] + \
+       (IAPU.DirectPage[OP1 + 1] << 8) + \
        IAPU.Registers.YA.B.Y;
 #endif
 
diff --git a/source/tile.c b/source/tile.c
index aa0cfaf..17a2934 100644
--- a/source/tile.c
+++ b/source/tile.c
@@ -779,9 +779,10 @@ static void WRITE_4PIXELS16_ADDF1_2(int32_t Offset, uint8_t* Pixels, uint16_t* S
    {
       if (GFX.Z1 > Depth [N] && (Pixel = Pixels[N]))
       {
-         Screen [N] = ScreenColors [Pixel];
          if (SubDepth [N] == 1)
             Screen [N] = (uint16_t)(COLOR_ADD1_2(ScreenColors [Pixel], GFX.FixedColour));
+         else
+            Screen [N] = ScreenColors [Pixel];
          Depth [N] = GFX.Z2;
       }
    }
@@ -798,9 +799,10 @@ static void WRITE_4PIXELS16_FLIPPED_ADDF1_2(int32_t Offset, uint8_t* Pixels, uin
    {
       if (GFX.Z1 > Depth [N] && (Pixel = Pixels[3 - N]))
       {
-         Screen [N] = ScreenColors [Pixel];
          if (SubDepth [N] == 1)
             Screen [N] = (uint16_t)(COLOR_ADD1_2(ScreenColors [Pixel], GFX.FixedColour));
+         else
+            Screen [N] = ScreenColors [Pixel];
          Depth [N] = GFX.Z2;
       }
    }
@@ -817,9 +819,10 @@ static void WRITE_4PIXELS16_SUBF1_2(int32_t Offset, uint8_t* Pixels, uint16_t* S
    {
       if (GFX.Z1 > Depth [N] && (Pixel = Pixels[N]))
       {
-         Screen [N] = ScreenColors [Pixel];
          if (SubDepth [N] == 1)
             Screen [N] = (uint16_t) COLOR_SUB1_2(ScreenColors [Pixel], GFX.FixedColour);
+         else
+            Screen [N] = ScreenColors [Pixel];
          Depth [N] = GFX.Z2;
       }
    }
@@ -836,9 +839,10 @@ static void WRITE_4PIXELS16_FLIPPED_SUBF1_2(int32_t Offset, uint8_t* Pixels, uin
    {
       if (GFX.Z1 > Depth [N] && (Pixel = Pixels[3 - N]))
       {
-         Screen [N] = ScreenColors [Pixel];
          if (SubDepth [N] == 1)
             Screen [N] = (uint16_t) COLOR_SUB1_2(ScreenColors [Pixel], GFX.FixedColour);
+         else
+            Screen [N] = ScreenColors [Pixel];
          Depth [N] = GFX.Z2;
       }
    }
@@ -924,7 +928,7 @@ void DrawLargePixel16Sub(uint32_t Tile, int32_t Offset, uint32_t StartPixel, uin
 
 #define LARGE_SUB_PIXEL(s, p) \
 (Depth [z + GFX.DepthDelta] ? (Depth [z + GFX.DepthDelta] != 1 ? \
-                COLOR_SUB (p, *(s + GFX.Delta))    : \
+                COLOR_SUB (p, *(s + GFX.Delta)) : \
                 COLOR_SUB (p, GFX.FixedColour)) : p)
 
    RENDER_TILE_LARGE(ScreenColors [pixel], LARGE_SUB_PIXEL);
@@ -940,7 +944,7 @@ void DrawLargePixel16Sub1_2(uint32_t Tile, int32_t Offset, uint32_t StartPixel,
 
 #define LARGE_SUB_PIXEL1_2(s, p) \
 (Depth [z + GFX.DepthDelta] ? (Depth [z + GFX.DepthDelta] != 1 ? \
-                COLOR_SUB1_2 (p, *(s + GFX.Delta))    : \
+                COLOR_SUB1_2 (p, *(s + GFX.Delta)) : \
                 COLOR_SUB (p, GFX.FixedColour)) : p)
 
    RENDER_TILE_LARGE(ScreenColors [pixel], LARGE_SUB_PIXEL1_2);