Un-inline a bunch of stuff.

With the MIPS instruction cache, this means that two consecutive SNES CPU instructions using e.g. the same addressing style or the same opcode have a chance that the second one will use the first one's code and that it will be cached.

10 files changed, 483 insertions, 406 deletions

diff --git a/Makefile b/Makefile
index a972f16..6bcaba6 100644
--- a/Makefile
+++ b/Makefile
@@ -63,7 +63,8 @@ OBJECTS      = $(C_OBJECTS) $(CPP_OBJECTS)
 # - - - Compilation flags - - -

 CFLAGS := -mips32 -Os -mno-abicalls -fno-pic -fno-builtin \

 	      -fno-exceptions -fno-function-sections -mlong-calls \

-	      -fomit-frame-pointer -msoft-float -G 4

+	      -fomit-frame-pointer -msoft-float -G 4 \

+          -fno-inline -fno-early-inlining

 

 DEFS   := -DSPC700_C -DEXECUTE_SUPERFX_PER_LINE -DSDD1_DECOMP \

           -DVAR_CYCLES -DCPU_SHUTDOWN -DSPC700_SHUTDOWN \

diff --git a/source/apumem.h b/source/apumem.h
index 8f6dfa4..573b52c 100644
--- a/source/apumem.h
+++ b/source/apumem.h
@@ -95,7 +95,7 @@ extern uint8 W4;
 extern uint8 APUROM[64];
 END_EXTERN_C
 
-INLINE uint8 S9xAPUGetByteZ (uint8 Address)
+uint8 S9xAPUGetByteZ (uint8 Address)
 {
     if (Address >= 0xf0 && IAPU.DirectPage == IAPU.RAM)
     {
@@ -127,7 +127,7 @@ INLINE uint8 S9xAPUGetByteZ (uint8 Address)
 	return (IAPU.DirectPage [Address]);
 }
 
-INLINE void S9xAPUSetByteZ (uint8 byte, uint8 Address)
+void S9xAPUSetByteZ (uint8 byte, uint8 Address)
 {
     if (Address >= 0xf0 && IAPU.DirectPage == IAPU.RAM)
     {
@@ -156,7 +156,7 @@ INLINE void S9xAPUSetByteZ (uint8 byte, uint8 Address)
 	IAPU.DirectPage [Address] = byte;
 }
 
-INLINE uint8 S9xAPUGetByte (uint32 Address)
+uint8 S9xAPUGetByte (uint32 Address)
 {
     Address &= 0xffff;
     
@@ -189,7 +189,7 @@ INLINE uint8 S9xAPUGetByte (uint32 Address)
 	return (IAPU.RAM [Address]);
 }
 
-INLINE void S9xAPUSetByte (uint8 byte, uint32 Address)
+void S9xAPUSetByte (uint8 byte, uint32 Address)
 {
     Address &= 0xffff;
     
diff --git a/source/cpuaddr.h b/source/cpuaddr.h
index a4c0ff4..76344db 100644
--- a/source/cpuaddr.h
+++ b/source/cpuaddr.h
@@ -100,19 +100,19 @@ typedef enum {
     JUMP = 4
 } AccessMode;
 
-STATIC inline void Immediate8 (AccessMode a)
+void Immediate8 (AccessMode a)
 {
     OpAddress = ICPU.ShiftedPB + CPU.PC - CPU.PCBase;
     CPU.PC++;
 }
 
-STATIC inline void Immediate16 (AccessMode a)
+void Immediate16 (AccessMode a)
 {
     OpAddress = ICPU.ShiftedPB + CPU.PC - CPU.PCBase;
     CPU.PC += 2;
 }
 
-STATIC inline void Relative (AccessMode a)
+void Relative (AccessMode a)
 {
     Int8 = *CPU.PC++;
 #ifndef SA1_OPCODES
@@ -121,7 +121,7 @@ STATIC inline void Relative (AccessMode a)
     OpAddress = ((int) (CPU.PC - CPU.PCBase) + Int8) & 0xffff;
 }
 
-STATIC inline void RelativeLong (AccessMode a)
+void RelativeLong (AccessMode a)
 {
 #ifdef FAST_LSB_WORD_ACCESS
     OpAddress = *(uint16 *) CPU.PC;
@@ -136,7 +136,7 @@ STATIC inline void RelativeLong (AccessMode a)
     OpAddress &= 0xffff;
 }
 
-STATIC inline void AbsoluteIndexedIndirect (AccessMode a)
+void AbsoluteIndexedIndirect (AccessMode a)
 {
 #ifdef FAST_LSB_WORD_ACCESS
     OpAddress = (Registers.X.W + *(uint16 *) CPU.PC) & 0xffff;
@@ -152,7 +152,7 @@ STATIC inline void AbsoluteIndexedIndirect (AccessMode a)
     if(a&READ) OpenBus = (uint8)(OpAddress>>8);
 }
 
-STATIC inline void AbsoluteIndirectLong (AccessMode a)
+void AbsoluteIndirectLong (AccessMode a)
 {
 #ifdef FAST_LSB_WORD_ACCESS
     OpAddress = *(uint16 *) CPU.PC;
@@ -172,7 +172,7 @@ STATIC inline void AbsoluteIndirectLong (AccessMode a)
     }
 }
 
-STATIC inline void AbsoluteIndirect (AccessMode a)
+void AbsoluteIndirect (AccessMode a)
 {
 #ifdef FAST_LSB_WORD_ACCESS
     OpAddress = *(uint16 *) CPU.PC;
@@ -190,7 +190,7 @@ STATIC inline void AbsoluteIndirect (AccessMode a)
     OpAddress += ICPU.ShiftedPB;
 }
 
-STATIC inline void Absolute (AccessMode a)
+void Absolute (AccessMode a)
 {
 #ifdef FAST_LSB_WORD_ACCESS
     OpAddress = *(uint16 *) CPU.PC + ICPU.ShiftedDB;
@@ -204,7 +204,7 @@ STATIC inline void Absolute (AccessMode a)
 #endif
 }
 
-STATIC inline void AbsoluteLong (AccessMode a)
+void AbsoluteLong (AccessMode a)
 {
 #ifdef FAST_LSB_WORD_ACCESS
     OpAddress = (*(uint32 *) CPU.PC) & 0xffffff;
@@ -218,7 +218,7 @@ STATIC inline void AbsoluteLong (AccessMode a)
 #endif
 }
 
-STATIC inline void Direct(AccessMode a)
+void Direct(AccessMode a)
 {
     if(a&READ) OpenBus = *CPU.PC;
     OpAddress = (*CPU.PC++ + Registers.D.W) & 0xffff;
@@ -228,7 +228,7 @@ STATIC inline void Direct(AccessMode a)
 //    if (Registers.DL != 0) CPU.Cycles += ONE_CYCLE;
 }
 
-STATIC inline void DirectIndirectIndexed (AccessMode a)
+void DirectIndirectIndexed (AccessMode a)
 {
     OpenBus = *CPU.PC;
     OpAddress = (*CPU.PC++ + Registers.D.W) & 0xffff;
@@ -245,7 +245,7 @@ STATIC inline void DirectIndirectIndexed (AccessMode a)
     // XXX: else Add one cycle if crosses page boundary
 }
 
-STATIC inline void DirectIndirectIndexedLong (AccessMode a)
+void DirectIndirectIndexedLong (AccessMode a)
 {
     OpenBus = *CPU.PC;
     OpAddress = (*CPU.PC++ + Registers.D.W) & 0xffff;
@@ -261,7 +261,7 @@ STATIC inline void DirectIndirectIndexedLong (AccessMode a)
 //    if (Registers.DL != 0) CPU.Cycles += ONE_CYCLE;
 }
 
-STATIC inline void DirectIndexedIndirect(AccessMode a)
+void DirectIndexedIndirect(AccessMode a)
 {
     OpenBus = *CPU.PC;
     OpAddress = (*CPU.PC++ + Registers.D.W + Registers.X.W) & 0xffff;
@@ -281,7 +281,7 @@ STATIC inline void DirectIndexedIndirect(AccessMode a)
 #endif
 }
 
-STATIC inline void DirectIndexedX (AccessMode a)
+void DirectIndexedX (AccessMode a)
 {
 	if(a&READ) OpenBus = *CPU.PC;
     OpAddress = (*CPU.PC++ + Registers.D.W + Registers.X.W);
@@ -299,7 +299,7 @@ STATIC inline void DirectIndexedX (AccessMode a)
 #endif
 }
 
-STATIC inline void DirectIndexedY (AccessMode a)
+void DirectIndexedY (AccessMode a)
 {
 	if(a&READ) OpenBus = *CPU.PC;
     OpAddress = (*CPU.PC++ + Registers.D.W + Registers.Y.W);
@@ -316,7 +316,7 @@ STATIC inline void DirectIndexedY (AccessMode a)
 #endif
 }
 
-STATIC inline void AbsoluteIndexedX (AccessMode a)
+void AbsoluteIndexedX (AccessMode a)
 {
 #ifdef FAST_LSB_WORD_ACCESS
     OpAddress = ICPU.ShiftedDB + *(uint16 *) CPU.PC + Registers.X.W;
@@ -333,7 +333,7 @@ STATIC inline void AbsoluteIndexedX (AccessMode a)
     // XXX: else is cross page boundary add one cycle
 }
 
-STATIC inline void AbsoluteIndexedY (AccessMode a)
+void AbsoluteIndexedY (AccessMode a)
 {
 #ifdef FAST_LSB_WORD_ACCESS
     OpAddress = ICPU.ShiftedDB + *(uint16 *) CPU.PC + Registers.Y.W;
@@ -350,7 +350,7 @@ STATIC inline void AbsoluteIndexedY (AccessMode a)
     // XXX: else is cross page boundary add one cycle
 }
 
-STATIC inline void AbsoluteLongIndexedX (AccessMode a)
+void AbsoluteLongIndexedX (AccessMode a)
 {
 #ifdef FAST_LSB_WORD_ACCESS
     OpAddress = (*(uint32 *) CPU.PC + Registers.X.W) & 0xffffff;
@@ -364,7 +364,7 @@ STATIC inline void AbsoluteLongIndexedX (AccessMode a)
 #endif
 }
 
-STATIC inline void DirectIndirect (AccessMode a)
+void DirectIndirect (AccessMode a)
 {
     OpenBus = *CPU.PC;
     OpAddress = (*CPU.PC++ + Registers.D.W) & 0xffff;
@@ -378,7 +378,7 @@ STATIC inline void DirectIndirect (AccessMode a)
 //    if (Registers.DL != 0) CPU.Cycles += ONE_CYCLE;
 }
 
-STATIC inline void DirectIndirectLong (AccessMode a)
+void DirectIndirectLong (AccessMode a)
 {
     OpenBus = *CPU.PC;
     OpAddress = (*CPU.PC++ + Registers.D.W) & 0xffff;
@@ -393,7 +393,7 @@ STATIC inline void DirectIndirectLong (AccessMode a)
 //    if (Registers.DL != 0) CPU.Cycles += ONE_CYCLE;
 }
 
-STATIC inline void StackRelative (AccessMode a)
+void StackRelative (AccessMode a)
 {
     if(a&READ) OpenBus = *CPU.PC;
     OpAddress = (*CPU.PC++ + Registers.S.W) & 0xffff;
@@ -403,7 +403,7 @@ STATIC inline void StackRelative (AccessMode a)
 #endif
 }
 
-STATIC inline void StackRelativeIndirectIndexed (AccessMode a)
+void StackRelativeIndirectIndexed (AccessMode a)
 {
     OpenBus = *CPU.PC;
     OpAddress = (*CPU.PC++ + Registers.S.W) & 0xffff;
diff --git a/source/cpumacro.h b/source/cpumacro.h
index d97aa53..5b4e15b 100644
--- a/source/cpumacro.h
+++ b/source/cpumacro.h
@@ -90,19 +90,19 @@
 #ifndef _CPUMACRO_H_
 #define _CPUMACRO_H_
 
-STATIC inline void SetZN16 (uint16 Work)
+void SetZN16 (uint16 Work)
 {
     ICPU._Zero = Work != 0;
     ICPU._Negative = (uint8) (Work >> 8);
 }
 
-STATIC inline void SetZN8 (uint8 Work)
+void SetZN8 (uint8 Work)
 {
     ICPU._Zero = Work;
     ICPU._Negative = Work;
 }
 
-STATIC inline void ADC8 ()
+void ADC8 ()
 {
     Work8 = S9xGetByte (OpAddress);
     
@@ -159,7 +159,7 @@ STATIC inline void ADC8 ()
     }
 }
 
-STATIC inline void ADC16 ()
+void ADC16 ()
 {
     Work16 = S9xGetWord (OpAddress);
 
@@ -235,19 +235,19 @@ STATIC inline void ADC16 ()
     }
 }
 
-STATIC inline void AND16 ()
+void AND16 ()
 {
     Registers.A.W &= S9xGetWord (OpAddress);
     SetZN16 (Registers.A.W);
 }
 
-STATIC inline void AND8 ()
+void AND8 ()
 {
     Registers.AL &= S9xGetByte (OpAddress);
     SetZN8 (Registers.AL);
 }
 
-STATIC inline void A_ASL16 ()
+void A_ASL16 ()
 {
 #ifndef SA1_OPCODES
     CPU.Cycles += ONE_CYCLE;
@@ -257,7 +257,7 @@ STATIC inline void A_ASL16 ()
     SetZN16 (Registers.A.W);
 }
 
-STATIC inline void A_ASL8 ()
+void A_ASL8 ()
 {
 #ifndef SA1_OPCODES
     CPU.Cycles += ONE_CYCLE;
@@ -267,7 +267,7 @@ STATIC inline void A_ASL8 ()
     SetZN8 (Registers.AL);
 }
 
-STATIC inline void ASL16 ()
+void ASL16 ()
 {
 #ifndef SA1_OPCODES
     CPU.Cycles += ONE_CYCLE;
@@ -281,7 +281,7 @@ STATIC inline void ASL16 ()
     SetZN16 (Work16);
 }
 
-STATIC inline void ASL8 ()
+void ASL8 ()
 {
 #ifndef SA1_OPCODES
     CPU.Cycles += ONE_CYCLE;
@@ -293,7 +293,7 @@ STATIC inline void ASL8 ()
     SetZN8 (Work8);
 }
 
-STATIC inline void BIT16 ()
+void BIT16 ()
 {
     Work16 = S9xGetWord (OpAddress);
     ICPU._Overflow = (Work16 & 0x4000) != 0;
@@ -301,7 +301,7 @@ STATIC inline void BIT16 ()
     ICPU._Zero = (Work16 & Registers.A.W) != 0;
 }
 
-STATIC inline void BIT8 ()
+void BIT8 ()
 {
     Work8 = S9xGetByte (OpAddress);
     ICPU._Overflow = (Work8 & 0x40) != 0;
@@ -309,7 +309,7 @@ STATIC inline void BIT8 ()
     ICPU._Zero = Work8 & Registers.AL;
 }
 
-STATIC inline void CMP16 ()
+void CMP16 ()
 {
     Int32 = (long) Registers.A.W -
 	    (long) S9xGetWord (OpAddress);
@@ -317,7 +317,7 @@ STATIC inline void CMP16 ()
     SetZN16 ((uint16) Int32);
 }
 
-STATIC inline void CMP8 ()
+void CMP8 ()
 {
     Int16 = (short) Registers.AL -
 	    (short) S9xGetByte (OpAddress);
@@ -325,7 +325,7 @@ STATIC inline void CMP8 ()
     SetZN8 ((uint8) Int16);
 }
 
-STATIC inline void CMX16 ()
+void CMX16 ()
 {
     Int32 = (long) Registers.X.W -
 	    (long) S9xGetWord (OpAddress);
@@ -333,7 +333,7 @@ STATIC inline void CMX16 ()
     SetZN16 ((uint16) Int32);
 }
 
-STATIC inline void CMX8 ()
+void CMX8 ()
 {
     Int16 = (short) Registers.XL -
 	    (short) S9xGetByte (OpAddress);
@@ -341,7 +341,7 @@ STATIC inline void CMX8 ()
     SetZN8 ((uint8) Int16);
 }
 
-STATIC inline void CMY16 ()
+void CMY16 ()
 {
     Int32 = (long) Registers.Y.W -
 	    (long) S9xGetWord (OpAddress);
@@ -349,7 +349,7 @@ STATIC inline void CMY16 ()
     SetZN16 ((uint16) Int32);
 }
 
-STATIC inline void CMY8 ()
+void CMY8 ()
 {
     Int16 = (short) Registers.YL -
 	    (short) S9xGetByte (OpAddress);
@@ -357,7 +357,7 @@ STATIC inline void CMY8 ()
     SetZN8 ((uint8) Int16);
 }
 
-STATIC inline void A_DEC16 ()
+void A_DEC16 ()
 {
 #ifndef SA1_OPCODES
     CPU.Cycles += ONE_CYCLE;
@@ -370,7 +370,7 @@ STATIC inline void A_DEC16 ()
     SetZN16 (Registers.A.W);
 }
 
-STATIC inline void A_DEC8 ()
+void A_DEC8 ()
 {
 #ifndef SA1_OPCODES
     CPU.Cycles += ONE_CYCLE;
@@ -383,7 +383,7 @@ STATIC inline void A_DEC8 ()
     SetZN8 (Registers.AL);
 }
 
-STATIC inline void DEC16 ()
+void DEC16 ()
 {
 #ifndef SA1_OPCODES
     CPU.Cycles += ONE_CYCLE;
@@ -399,7 +399,7 @@ STATIC inline void DEC16 ()
 	SetZN16 (Work16);
 }
 
-STATIC inline void DEC8 ()
+void DEC8 ()
 {
 #ifndef SA1_OPCODES
     CPU.Cycles += ONE_CYCLE;
@@ -413,19 +413,19 @@ STATIC inline void DEC8 ()
     SetZN8 (Work8);
 }
 
-STATIC inline void EOR16 ()
+void EOR16 ()
 {
     Registers.A.W ^= S9xGetWord (OpAddress);
     SetZN16 (Registers.A.W);
 }
 
-STATIC inline void EOR8 ()
+void EOR8 ()
 {
     Registers.AL ^= S9xGetByte (OpAddress);
     SetZN8 (Registers.AL);
 }
 
-STATIC inline void A_INC16 ()
+void A_INC16 ()
 {
 #ifndef SA1_OPCODES
     CPU.Cycles += ONE_CYCLE;
@@ -438,7 +438,7 @@ STATIC inline void A_INC16 ()
     SetZN16 (Registers.A.W);
 }
 
-STATIC inline void A_INC8 ()
+void A_INC8 ()
 {
 #ifndef SA1_OPCODES
     CPU.Cycles += ONE_CYCLE;
@@ -451,7 +451,7 @@ STATIC inline void A_INC8 ()
     SetZN8 (Registers.AL);
 }
 
-STATIC inline void INC16 ()
+void INC16 ()
 {
 #ifndef SA1_OPCODES
     CPU.Cycles += ONE_CYCLE;
@@ -467,7 +467,7 @@ STATIC inline void INC16 ()
     SetZN16 (Work16);
 }
 
-STATIC inline void INC8 ()
+void INC8 ()
 {
 #ifndef SA1_OPCODES
     CPU.Cycles += ONE_CYCLE;
@@ -481,43 +481,43 @@ STATIC inline void INC8 ()
     SetZN8 (Work8);
 }
 
-STATIC inline void LDA16 ()
+void LDA16 ()
 {
     Registers.A.W = S9xGetWord (OpAddress);
     SetZN16 (Registers.A.W);
 }
 
-STATIC inline void LDA8 ()
+void LDA8 ()
 {
     Registers.AL = S9xGetByte (OpAddress);
     SetZN8 (Registers.AL);
 }
 
-STATIC inline void LDX16 ()
+void LDX16 ()
 {
     Registers.X.W = S9xGetWord (OpAddress);
     SetZN16 (Registers.X.W);
 }
 
-STATIC inline void LDX8 ()
+void LDX8 ()
 {
     Registers.XL = S9xGetByte (OpAddress);
     SetZN8 (Registers.XL);
 }
 
-STATIC inline void LDY16 ()
+void LDY16 ()
 {
     Registers.Y.W = S9xGetWord (OpAddress);
     SetZN16 (Registers.Y.W);
 }
 
-STATIC inline void LDY8 ()
+void LDY8 ()
 {
     Registers.YL = S9xGetByte (OpAddress);
     SetZN8 (Registers.YL);
 }
 
-STATIC inline void A_LSR16 ()
+void A_LSR16 ()
 {
 #ifndef SA1_OPCODES
     CPU.Cycles += ONE_CYCLE;
@@ -527,7 +527,7 @@ STATIC inline void A_LSR16 ()
     SetZN16 (Registers.A.W);
 }
 
-STATIC inline void A_LSR8 ()
+void A_LSR8 ()
 {
 #ifndef SA1_OPCODES
     CPU.Cycles += ONE_CYCLE;
@@ -537,7 +537,7 @@ STATIC inline void A_LSR8 ()
     SetZN8 (Registers.AL);
 }
 
-STATIC inline void LSR16 ()
+void LSR16 ()
 {
 #ifndef SA1_OPCODES
     CPU.Cycles += ONE_CYCLE;
@@ -551,7 +551,7 @@ STATIC inline void LSR16 ()
     SetZN16 (Work16);
 }
 
-STATIC inline void LSR8 ()
+void LSR8 ()
 {
 #ifndef SA1_OPCODES
     CPU.Cycles += ONE_CYCLE;
@@ -563,19 +563,19 @@ STATIC inline void LSR8 ()
     SetZN8 (Work8);
 }
 
-STATIC inline void ORA16 ()
+void ORA16 ()
 {
     Registers.A.W |= S9xGetWord (OpAddress);
     SetZN16 (Registers.A.W);
 }
 
-STATIC inline void ORA8 ()
+void ORA8 ()
 {
     Registers.AL |= S9xGetByte (OpAddress);
     SetZN8 (Registers.AL);
 }
 
-STATIC inline void A_ROL16 ()
+void A_ROL16 ()
 {
 #ifndef SA1_OPCODES
     CPU.Cycles += ONE_CYCLE;
@@ -586,7 +586,7 @@ STATIC inline void A_ROL16 ()
     SetZN16 ((uint16) Work32);
 }
 
-STATIC inline void A_ROL8 ()
+void A_ROL8 ()
 {
 #ifndef SA1_OPCODES
     CPU.Cycles += ONE_CYCLE;
@@ -599,7 +599,7 @@ STATIC inline void A_ROL8 ()
     SetZN8 ((uint8) Work16);
 }
 
-STATIC inline void ROL16 ()
+void ROL16 ()
 {
 #ifndef SA1_OPCODES
     CPU.Cycles += ONE_CYCLE;
@@ -614,7 +614,7 @@ STATIC inline void ROL16 ()
     SetZN16 ((uint16) Work32);
 }
 
-STATIC inline void ROL8 ()
+void ROL8 ()
 {
 #ifndef SA1_OPCODES
     CPU.Cycles += ONE_CYCLE;
@@ -627,7 +627,7 @@ STATIC inline void ROL8 ()
     SetZN8 ((uint8) Work16);
 }
 
-STATIC inline void A_ROR16 ()
+void A_ROR16 ()
 {
 #ifndef SA1_OPCODES
     CPU.Cycles += ONE_CYCLE;
@@ -640,7 +640,7 @@ STATIC inline void A_ROR16 ()
     SetZN16 ((uint16) Work32);
 }
 
-STATIC inline void A_ROR8 ()
+void A_ROR8 ()
 {
 #ifndef SA1_OPCODES
     CPU.Cycles += ONE_CYCLE;
@@ -652,7 +652,7 @@ STATIC inline void A_ROR8 ()
     SetZN8 ((uint8) Work16);
 }
 
-STATIC inline void ROR16 ()
+void ROR16 ()
 {
 #ifndef SA1_OPCODES
     CPU.Cycles += ONE_CYCLE;
@@ -667,7 +667,7 @@ STATIC inline void ROR16 ()
     SetZN16 ((uint16) Work32);
 }
 
-STATIC inline void ROR8 ()
+void ROR8 ()
 {
 #ifndef SA1_OPCODES
     CPU.Cycles += ONE_CYCLE;
@@ -680,7 +680,7 @@ STATIC inline void ROR8 ()
     SetZN8 ((uint8) Work16);
 }
 
-STATIC inline void SBC16 ()
+void SBC16 ()
 {
     Work16 = S9xGetWord (OpAddress);
 
@@ -750,7 +750,7 @@ STATIC inline void SBC16 ()
     }
 }
 
-STATIC inline void SBC8 ()
+void SBC8 ()
 {
     Work8 = S9xGetByte (OpAddress);
     if (CheckDecimal ())
@@ -801,47 +801,47 @@ STATIC inline void SBC8 ()
     }
 }
 
-STATIC inline void STA16 ()
+void STA16 ()
 {
     S9xSetWord (Registers.A.W, OpAddress);
 }
 
-STATIC inline void STA8 ()
+void STA8 ()
 {
     S9xSetByte (Registers.AL, OpAddress);
 }
 
-STATIC inline void STX16 ()
+void STX16 ()
 {
     S9xSetWord (Registers.X.W, OpAddress);
 }
 
-STATIC inline void STX8 ()
+void STX8 ()
 {
     S9xSetByte (Registers.XL, OpAddress);
 }
 
-STATIC inline void STY16 ()
+void STY16 ()
 {
     S9xSetWord (Registers.Y.W, OpAddress);
 }
 
-STATIC inline void STY8 ()
+void STY8 ()
 {
     S9xSetByte (Registers.YL, OpAddress);
 }
 
-STATIC inline void STZ16 ()
+void STZ16 ()
 {
     S9xSetWord (0, OpAddress);
 }
 
-STATIC inline void STZ8 ()
+void STZ8 ()
 {
     S9xSetByte (0, OpAddress);
 }
 
-STATIC inline void TSB16 ()
+void TSB16 ()
 {
 #ifndef SA1_OPCODES
     CPU.Cycles += ONE_CYCLE;
@@ -854,7 +854,7 @@ STATIC inline void TSB16 ()
 	S9xSetByte (Work16&0xFF, OpAddress);
 }
 
-STATIC inline void TSB8 ()
+void TSB8 ()
 {
 #ifndef SA1_OPCODES
     CPU.Cycles += ONE_CYCLE;
@@ -865,7 +865,7 @@ STATIC inline void TSB8 ()
     S9xSetByte (Work8, OpAddress);
 }
 
-STATIC inline void TRB16 ()
+void TRB16 ()
 {
 #ifndef SA1_OPCODES
     CPU.Cycles += ONE_CYCLE;
@@ -878,7 +878,7 @@ STATIC inline void TRB16 ()
 	S9xSetByte (Work16&0xFF, OpAddress);
 }
 
-STATIC inline void TRB8 ()
+void TRB8 ()
 {
 #ifndef SA1_OPCODES
     CPU.Cycles += ONE_CYCLE;
diff --git a/source/getset.h b/source/getset.h
index 2368ac2..f04bf95 100644
--- a/source/getset.h
+++ b/source/getset.h
@@ -103,7 +103,7 @@ extern "C"
 	extern uint8 OpenBus;
 }
 
-INLINE uint8 S9xGetByte (uint32 Address)
+uint8 S9xGetByte (uint32 Address)
 {
     int block;
     uint8 *GetAddress = Memory.Map [block = (Address >> MEMMAP_SHIFT) & MEMMAP_MASK];
@@ -197,7 +197,7 @@ INLINE uint8 S9xGetByte (uint32 Address)
     }
 }
 
-INLINE uint16 S9xGetWord (uint32 Address)
+uint16 S9xGetWord (uint32 Address)
 {
     if ((Address & 0x0fff) == 0x0fff)
     {
@@ -320,7 +320,7 @@ INLINE uint16 S9xGetWord (uint32 Address)
     }
 }
 
-INLINE void S9xSetByte (uint8 Byte, uint32 Address)
+void S9xSetByte (uint8 Byte, uint32 Address)
 {
 #if defined(CPU_SHUTDOWN)
     CPU.WaitAddress = NULL;
@@ -436,7 +436,7 @@ INLINE void S9xSetByte (uint8 Byte, uint32 Address)
     }
 }
 
-INLINE void S9xSetWord (uint16 Word, uint32 Address)
+void S9xSetWord (uint16 Word, uint32 Address)
 {
 	if((Address & 0x0FFF)==0x0FFF)
 	{
@@ -594,7 +594,7 @@ INLINE void S9xSetWord (uint16 Word, uint32 Address)
     }
 }
 
-INLINE uint8 *GetBasePointer (uint32 Address)
+uint8 *GetBasePointer (uint32 Address)
 {
     uint8 *GetAddress = Memory.Map [(Address >> MEMMAP_SHIFT) & MEMMAP_MASK];
     if (GetAddress >= (uint8 *) CMemory::MAP_LAST)
@@ -660,7 +660,7 @@ INLINE uint8 *GetBasePointer (uint32 Address)
     }
 }
 
-INLINE uint8 *S9xGetMemPointer (uint32 Address)
+uint8 *S9xGetMemPointer (uint32 Address)
 {
     uint8 *GetAddress = Memory.Map [(Address >> MEMMAP_SHIFT) & MEMMAP_MASK];
     if (GetAddress >= (uint8 *) CMemory::MAP_LAST)
@@ -714,7 +714,7 @@ INLINE uint8 *S9xGetMemPointer (uint32 Address)
     }
 }
 
-INLINE void S9xSetPCBase (uint32 Address)
+void S9xSetPCBase (uint32 Address)
 {
     int block;
     uint8 *GetAddress = Memory.Map [block = (Address >> MEMMAP_SHIFT) & MEMMAP_MASK];
diff --git a/source/memmap.cpp b/source/memmap.cpp
index 65b8739..e378601 100644
--- a/source/memmap.cpp
+++ b/source/memmap.cpp
@@ -146,7 +146,7 @@ int is_bsx(unsigned char *);
 int bs_name(unsigned char *);
 int check_char(unsigned);
 void S9xDeinterleaveType2 (bool8 reset=TRUE);
-inline uint32 caCRC32(uint8 *array, uint32 size, register uint32 crc32 = 0xFFFFFFFF);
+uint32 caCRC32(uint8 *array, uint32 size, register uint32 crc32 = 0xFFFFFFFF);
 
 extern char *rom_filename;
 
@@ -395,10 +395,11 @@ char *CMemory::Safe (const char *s)
 
 /**********************************************************************************************/
 /* Init()                                                                                     */
-/* This function allocates all the memory needed by the emulator                              */
+/* This function allocates and zeroes all the memory needed by the emulator                   */
 /**********************************************************************************************/
 bool8 CMemory::Init ()
 {
+	// memset? Really? We could just memcpy after the first allocation... [Neb]
     RAM	    = (uint8 *) malloc (0x20000);
     SRAM    = (uint8 *) malloc (0x20000);
     VRAM    = (uint8 *) malloc (0x10000);
@@ -1213,7 +1214,7 @@ void S9xDeinterleaveType2 (bool8 reset)
 }
 
 //CRC32 for char arrays
-inline uint32 caCRC32(uint8 *array, uint32 size, register uint32 crc32)
+uint32 caCRC32(uint8 *array, uint32 size, register uint32 crc32)
 {
   for (register uint32 i = 0; i < size; i++)
   {
diff --git a/source/ppu.cpp b/source/ppu.cpp
index fc0bd51..389bdd2 100644
--- a/source/ppu.cpp
+++ b/source/ppu.cpp
@@ -117,7 +117,7 @@ uint8 in_bit=0;
 
 extern uint8 *HDMAMemPointers [8];
 
-static inline void S9xLatchCounters (bool force)
+void S9xLatchCounters (bool force)
 {
 	if(!force && !(Memory.FillRAM[0x4213] & 0x80)) return;
 
@@ -3297,3 +3297,286 @@ printf ("%06x: %d\n", t, FxEmulate (2000000));
 }
 #endif
 
+// Register reads and writes...
+
+uint8 REGISTER_4212()
+{
+    GetBank = 0;
+    if (CPU.V_Counter >= PPU.ScreenHeight + FIRST_VISIBLE_LINE &&
+    CPU.V_Counter < PPU.ScreenHeight + FIRST_VISIBLE_LINE + 3)
+    GetBank = 1;
+
+    GetBank |= CPU.Cycles >= Settings.HBlankStart ? 0x40 : 0;
+    if (CPU.V_Counter >= PPU.ScreenHeight + FIRST_VISIBLE_LINE)
+    GetBank |= 0x80; /* XXX: 0x80 or 0xc0 ? */
+
+    return (GetBank);
+}
+
+void FLUSH_REDRAW ()
+{
+    if (IPPU.PreviousLine != IPPU.CurrentLine)
+    S9xUpdateScreen ();
+}
+
+void REGISTER_2104 (uint8 byte)
+{
+    if (PPU.OAMAddr & 0x100)
+    {
+        int addr = ((PPU.OAMAddr & 0x10f) << 1) + (PPU.OAMFlip & 1);
+        if (byte != PPU.OAMData [addr]){
+            FLUSH_REDRAW ();
+            PPU.OAMData [addr] = byte;
+            IPPU.OBJChanged = TRUE;
+
+            // X position high bit, and sprite size (x4)
+            struct SOBJ *pObj = &PPU.OBJ [(addr & 0x1f) * 4];
+
+            pObj->HPos = (pObj->HPos & 0xFF) | SignExtend[(byte >> 0) & 1];
+            pObj++->Size = byte & 2;
+            pObj->HPos = (pObj->HPos & 0xFF) | SignExtend[(byte >> 2) & 1];
+            pObj++->Size = byte & 8;
+            pObj->HPos = (pObj->HPos & 0xFF) | SignExtend[(byte >> 4) & 1];
+            pObj++->Size = byte & 32;
+            pObj->HPos = (pObj->HPos & 0xFF) | SignExtend[(byte >> 6) & 1];
+            pObj->Size = byte & 128;
+        }
+        PPU.OAMFlip ^= 1;
+        if(!(PPU.OAMFlip & 1)){
+            ++PPU.OAMAddr;
+            PPU.OAMAddr &= 0x1ff;
+            if (PPU.OAMPriorityRotation && PPU.FirstSprite != (PPU.OAMAddr >> 1))
+            {
+                PPU.FirstSprite = (PPU.OAMAddr&0xFE) >> 1;
+                IPPU.OBJChanged = TRUE;
+            }
+        } else {
+            if (PPU.OAMPriorityRotation && (PPU.OAMAddr&1)) IPPU.OBJChanged = TRUE;
+        }
+    } else if(!(PPU.OAMFlip & 1)){
+        PPU.OAMWriteRegister &= 0xff00;
+        PPU.OAMWriteRegister |= byte;
+        PPU.OAMFlip |= 1;
+        if (PPU.OAMPriorityRotation && (PPU.OAMAddr&1)) IPPU.OBJChanged = TRUE;
+    } else {
+        PPU.OAMWriteRegister &= 0x00ff;
+        uint8 lowbyte = (uint8)(PPU.OAMWriteRegister);
+        uint8 highbyte = byte;
+        PPU.OAMWriteRegister |= byte << 8;
+
+        int addr = (PPU.OAMAddr << 1);
+
+        if (lowbyte != PPU.OAMData [addr] ||
+            highbyte != PPU.OAMData [addr+1])
+        {
+            FLUSH_REDRAW ();
+            PPU.OAMData [addr] = lowbyte;
+            PPU.OAMData [addr+1] = highbyte;
+            IPPU.OBJChanged = TRUE;
+            if (addr & 2)
+            {
+                // Tile
+                PPU.OBJ[addr = PPU.OAMAddr >> 1].Name = PPU.OAMWriteRegister & 0x1ff;
+
+                // priority, h and v flip.
+                PPU.OBJ[addr].Palette = (highbyte >> 1) & 7;
+                PPU.OBJ[addr].Priority = (highbyte >> 4) & 3;
+                PPU.OBJ[addr].HFlip = (highbyte >> 6) & 1;
+                PPU.OBJ[addr].VFlip = (highbyte >> 7) & 1;
+            }
+            else
+            {
+                // X position (low)
+                PPU.OBJ[addr = PPU.OAMAddr >> 1].HPos &= 0xFF00;
+                PPU.OBJ[addr].HPos |= lowbyte;
+
+                // Sprite Y position
+                PPU.OBJ[addr].VPos = highbyte;
+            }
+        }
+        PPU.OAMFlip &= ~1;
+        ++PPU.OAMAddr;
+        if (PPU.OAMPriorityRotation && PPU.FirstSprite != (PPU.OAMAddr >> 1))
+        {
+            PPU.FirstSprite = (PPU.OAMAddr&0xFE) >> 1;
+            IPPU.OBJChanged = TRUE;
+        }
+    }
+
+    Memory.FillRAM [0x2104] = byte;
+}
+
+void REGISTER_2118 (uint8 Byte)
+{
+    uint32 address;
+    if (PPU.VMA.FullGraphicCount)
+    {
+    uint32 rem = PPU.VMA.Address & PPU.VMA.Mask1;
+    address = (((PPU.VMA.Address & ~PPU.VMA.Mask1) +
+             (rem >> PPU.VMA.Shift) +
+             ((rem & (PPU.VMA.FullGraphicCount - 1)) << 3)) << 1) & 0xffff;
+    Memory.VRAM [address] = Byte;
+    }
+    else
+    {
+    Memory.VRAM[address = (PPU.VMA.Address << 1) & 0xFFFF] = Byte;
+    }
+    IPPU.TileCached [TILE_2BIT][address >> 4] = FALSE;
+    IPPU.TileCached [TILE_4BIT][address >> 5] = FALSE;
+    IPPU.TileCached [TILE_8BIT][address >> 6] = FALSE;
+    if (!PPU.VMA.High)
+    {
+#ifdef DEBUGGER
+    if (Settings.TraceVRAM && !CPU.InDMA)
+    {
+        printf ("VRAM write byte: $%04X (%d,%d)\n", PPU.VMA.Address,
+            Memory.FillRAM[0x2115] & 3,
+            (Memory.FillRAM [0x2115] & 0x0c) >> 2);
+    }
+#endif  
+    PPU.VMA.Address += PPU.VMA.Increment;
+    }
+//    Memory.FillRAM [0x2118] = Byte;
+}
+
+void REGISTER_2118_tile (uint8 Byte)
+{
+    uint32 address;
+    uint32 rem = PPU.VMA.Address & PPU.VMA.Mask1;
+    address = (((PPU.VMA.Address & ~PPU.VMA.Mask1) +
+         (rem >> PPU.VMA.Shift) +
+         ((rem & (PPU.VMA.FullGraphicCount - 1)) << 3)) << 1) & 0xffff;
+    Memory.VRAM [address] = Byte;
+    IPPU.TileCached [TILE_2BIT][address >> 4] = FALSE;
+    IPPU.TileCached [TILE_4BIT][address >> 5] = FALSE;
+    IPPU.TileCached [TILE_8BIT][address >> 6] = FALSE;
+    if (!PPU.VMA.High)
+    PPU.VMA.Address += PPU.VMA.Increment;
+//    Memory.FillRAM [0x2118] = Byte;
+}
+
+void REGISTER_2118_linear (uint8 Byte)
+{
+    uint32 address;
+    Memory.VRAM[address = (PPU.VMA.Address << 1) & 0xFFFF] = Byte;
+    IPPU.TileCached [TILE_2BIT][address >> 4] = FALSE;
+    IPPU.TileCached [TILE_4BIT][address >> 5] = FALSE;
+    IPPU.TileCached [TILE_8BIT][address >> 6] = FALSE;
+    if (!PPU.VMA.High)
+    PPU.VMA.Address += PPU.VMA.Increment;
+//    Memory.FillRAM [0x2118] = Byte;
+}
+
+void REGISTER_2119 (uint8 Byte)
+{
+    uint32 address;
+    if (PPU.VMA.FullGraphicCount)
+    {
+    uint32 rem = PPU.VMA.Address & PPU.VMA.Mask1;
+    address = ((((PPU.VMA.Address & ~PPU.VMA.Mask1) +
+            (rem >> PPU.VMA.Shift) +
+            ((rem & (PPU.VMA.FullGraphicCount - 1)) << 3)) << 1) + 1) & 0xFFFF;
+    Memory.VRAM [address] = Byte;
+    }
+    else
+    {
+    Memory.VRAM[address = ((PPU.VMA.Address << 1) + 1) & 0xFFFF] = Byte;
+    }
+    IPPU.TileCached [TILE_2BIT][address >> 4] = FALSE;
+    IPPU.TileCached [TILE_4BIT][address >> 5] = FALSE;
+    IPPU.TileCached [TILE_8BIT][address >> 6] = FALSE;
+    if (PPU.VMA.High)
+    {
+#ifdef DEBUGGER
+    if (Settings.TraceVRAM && !CPU.InDMA)
+    {
+        printf ("VRAM write word: $%04X (%d,%d)\n", PPU.VMA.Address,
+            Memory.FillRAM[0x2115] & 3,
+            (Memory.FillRAM [0x2115] & 0x0c) >> 2);
+    }
+#endif  
+    PPU.VMA.Address += PPU.VMA.Increment;
+    }
+//    Memory.FillRAM [0x2119] = Byte;
+}
+
+void REGISTER_2119_tile (uint8 Byte)
+{
+    uint32 rem = PPU.VMA.Address & PPU.VMA.Mask1;
+    uint32 address = ((((PPU.VMA.Address & ~PPU.VMA.Mask1) +
+            (rem >> PPU.VMA.Shift) +
+            ((rem & (PPU.VMA.FullGraphicCount - 1)) << 3)) << 1) + 1) & 0xFFFF;
+    Memory.VRAM [address] = Byte;
+    IPPU.TileCached [TILE_2BIT][address >> 4] = FALSE;
+    IPPU.TileCached [TILE_4BIT][address >> 5] = FALSE;
+    IPPU.TileCached [TILE_8BIT][address >> 6] = FALSE;
+    if (PPU.VMA.High)
+    PPU.VMA.Address += PPU.VMA.Increment;
+//    Memory.FillRAM [0x2119] = Byte;
+}
+
+void REGISTER_2119_linear (uint8 Byte)
+{
+    uint32 address;
+    Memory.VRAM[address = ((PPU.VMA.Address << 1) + 1) & 0xFFFF] = Byte;
+    IPPU.TileCached [TILE_2BIT][address >> 4] = FALSE;
+    IPPU.TileCached [TILE_4BIT][address >> 5] = FALSE;
+    IPPU.TileCached [TILE_8BIT][address >> 6] = FALSE;
+    if (PPU.VMA.High)
+    PPU.VMA.Address += PPU.VMA.Increment;
+//    Memory.FillRAM [0x2119] = Byte;
+}
+
+void REGISTER_2122(uint8 Byte)
+{
+    // CG-RAM (palette) write
+
+    if (PPU.CGFLIP)
+    {
+    if ((Byte & 0x7f) != (PPU.CGDATA[PPU.CGADD] >> 8))
+    {
+        if (Settings.SixteenBit)
+        FLUSH_REDRAW ();
+        PPU.CGDATA[PPU.CGADD] &= 0x00FF;
+        PPU.CGDATA[PPU.CGADD] |= (Byte & 0x7f) << 8;
+        IPPU.ColorsChanged = TRUE;
+        if (Settings.SixteenBit)
+        {
+        IPPU.Blue [PPU.CGADD] = IPPU.XB [(Byte >> 2) & 0x1f];
+        IPPU.Green [PPU.CGADD] = IPPU.XB [(PPU.CGDATA[PPU.CGADD] >> 5) & 0x1f];
+        IPPU.ScreenColors [PPU.CGADD] = (uint16) BUILD_PIXEL (IPPU.Red [PPU.CGADD],
+                                 IPPU.Green [PPU.CGADD],
+                                 IPPU.Blue [PPU.CGADD]);
+        }
+    }
+    PPU.CGADD++;
+    }
+    else
+    {
+    if (Byte != (uint8) (PPU.CGDATA[PPU.CGADD] & 0xff))
+    {
+        if (Settings.SixteenBit)
+        FLUSH_REDRAW ();
+        PPU.CGDATA[PPU.CGADD] &= 0x7F00;
+        PPU.CGDATA[PPU.CGADD] |= Byte;
+        IPPU.ColorsChanged = TRUE;
+        if (Settings.SixteenBit)
+        {
+        IPPU.Red [PPU.CGADD] = IPPU.XB [Byte & 0x1f];
+        IPPU.Green [PPU.CGADD] = IPPU.XB [(PPU.CGDATA[PPU.CGADD] >> 5) & 0x1f];
+        IPPU.ScreenColors [PPU.CGADD] = (uint16) BUILD_PIXEL (IPPU.Red [PPU.CGADD],
+                                 IPPU.Green [PPU.CGADD],
+                                 IPPU.Blue [PPU.CGADD]);
+        }
+    }
+    }
+    PPU.CGFLIP ^= 1;
+//    Memory.FillRAM [0x2122] = Byte;
+}
+
+void REGISTER_2180(uint8 Byte)
+{
+    Memory.RAM[PPU.WRAM++] = Byte;
+    PPU.WRAM &= 0x1FFFF;
+    Memory.FillRAM [0x2180] = Byte;
+}
diff --git a/source/ppu.h b/source/ppu.h
index 99dde34..5a889ef 100644
--- a/source/ppu.h
+++ b/source/ppu.h
@@ -342,288 +342,17 @@ END_EXTERN_C
 #define MAX_5C78_VERSION 0x03
 #define MAX_5A22_VERSION 0x02
 
-STATIC inline uint8 REGISTER_4212()
-{
-    GetBank = 0;
-    if (CPU.V_Counter >= PPU.ScreenHeight + FIRST_VISIBLE_LINE &&
-	CPU.V_Counter < PPU.ScreenHeight + FIRST_VISIBLE_LINE + 3)
-	GetBank = 1;
-
-    GetBank |= CPU.Cycles >= Settings.HBlankStart ? 0x40 : 0;
-    if (CPU.V_Counter >= PPU.ScreenHeight + FIRST_VISIBLE_LINE)
-	GetBank |= 0x80; /* XXX: 0x80 or 0xc0 ? */
-
-    return (GetBank);
-}
-
-STATIC inline void FLUSH_REDRAW ()
-{
-    if (IPPU.PreviousLine != IPPU.CurrentLine)
-	S9xUpdateScreen ();
-}
-
-STATIC inline void REGISTER_2104 (uint8 byte)
-{
-    if (PPU.OAMAddr & 0x100)
-    {
-        int addr = ((PPU.OAMAddr & 0x10f) << 1) + (PPU.OAMFlip & 1);
-        if (byte != PPU.OAMData [addr]){
-            FLUSH_REDRAW ();
-            PPU.OAMData [addr] = byte;
-            IPPU.OBJChanged = TRUE;
-
-            // X position high bit, and sprite size (x4)
-            struct SOBJ *pObj = &PPU.OBJ [(addr & 0x1f) * 4];
-
-            pObj->HPos = (pObj->HPos & 0xFF) | SignExtend[(byte >> 0) & 1];
-            pObj++->Size = byte & 2;
-            pObj->HPos = (pObj->HPos & 0xFF) | SignExtend[(byte >> 2) & 1];
-            pObj++->Size = byte & 8;
-            pObj->HPos = (pObj->HPos & 0xFF) | SignExtend[(byte >> 4) & 1];
-            pObj++->Size = byte & 32;
-            pObj->HPos = (pObj->HPos & 0xFF) | SignExtend[(byte >> 6) & 1];
-            pObj->Size = byte & 128;
-        }
-        PPU.OAMFlip ^= 1;
-		if(!(PPU.OAMFlip & 1)){
-			++PPU.OAMAddr;
-			PPU.OAMAddr &= 0x1ff;
-			if (PPU.OAMPriorityRotation && PPU.FirstSprite != (PPU.OAMAddr >> 1))
-			{
-				PPU.FirstSprite = (PPU.OAMAddr&0xFE) >> 1;
-				IPPU.OBJChanged = TRUE;
-			}
-		} else {
-			if (PPU.OAMPriorityRotation && (PPU.OAMAddr&1)) IPPU.OBJChanged = TRUE;
-		}
-    } else if(!(PPU.OAMFlip & 1)){
-        PPU.OAMWriteRegister &= 0xff00;
-        PPU.OAMWriteRegister |= byte;
-        PPU.OAMFlip |= 1;
-		if (PPU.OAMPriorityRotation && (PPU.OAMAddr&1)) IPPU.OBJChanged = TRUE;
-    } else {
-        PPU.OAMWriteRegister &= 0x00ff;
-        uint8 lowbyte = (uint8)(PPU.OAMWriteRegister);
-        uint8 highbyte = byte;
-        PPU.OAMWriteRegister |= byte << 8;
-
-        int addr = (PPU.OAMAddr << 1);
-
-        if (lowbyte != PPU.OAMData [addr] ||
-            highbyte != PPU.OAMData [addr+1])
-        {
-            FLUSH_REDRAW ();
-            PPU.OAMData [addr] = lowbyte;
-            PPU.OAMData [addr+1] = highbyte;
-            IPPU.OBJChanged = TRUE;
-            if (addr & 2)
-            {
-                // Tile
-                PPU.OBJ[addr = PPU.OAMAddr >> 1].Name = PPU.OAMWriteRegister & 0x1ff;
-
-                // priority, h and v flip.
-                PPU.OBJ[addr].Palette = (highbyte >> 1) & 7;
-                PPU.OBJ[addr].Priority = (highbyte >> 4) & 3;
-                PPU.OBJ[addr].HFlip = (highbyte >> 6) & 1;
-                PPU.OBJ[addr].VFlip = (highbyte >> 7) & 1;
-            }
-            else
-            {
-                // X position (low)
-                PPU.OBJ[addr = PPU.OAMAddr >> 1].HPos &= 0xFF00;
-                PPU.OBJ[addr].HPos |= lowbyte;
-
-                // Sprite Y position
-                PPU.OBJ[addr].VPos = highbyte;
-            }
-        }
-        PPU.OAMFlip &= ~1;
-        ++PPU.OAMAddr;
-		if (PPU.OAMPriorityRotation && PPU.FirstSprite != (PPU.OAMAddr >> 1))
-		{
-			PPU.FirstSprite = (PPU.OAMAddr&0xFE) >> 1;
-			IPPU.OBJChanged = TRUE;
-		}
-    }
-
-    Memory.FillRAM [0x2104] = byte;
-}
-
-STATIC inline void REGISTER_2118 (uint8 Byte)
-{
-    uint32 address;
-    if (PPU.VMA.FullGraphicCount)
-    {
-	uint32 rem = PPU.VMA.Address & PPU.VMA.Mask1;
-	address = (((PPU.VMA.Address & ~PPU.VMA.Mask1) +
-			 (rem >> PPU.VMA.Shift) +
-			 ((rem & (PPU.VMA.FullGraphicCount - 1)) << 3)) << 1) & 0xffff;
-	Memory.VRAM [address] = Byte;
-    }
-    else
-    {
-	Memory.VRAM[address = (PPU.VMA.Address << 1) & 0xFFFF] = Byte;
-    }
-    IPPU.TileCached [TILE_2BIT][address >> 4] = FALSE;
-    IPPU.TileCached [TILE_4BIT][address >> 5] = FALSE;
-    IPPU.TileCached [TILE_8BIT][address >> 6] = FALSE;
-    if (!PPU.VMA.High)
-    {
-#ifdef DEBUGGER
-	if (Settings.TraceVRAM && !CPU.InDMA)
-	{
-	    printf ("VRAM write byte: $%04X (%d,%d)\n", PPU.VMA.Address,
-		    Memory.FillRAM[0x2115] & 3,
-		    (Memory.FillRAM [0x2115] & 0x0c) >> 2);
-	}
-#endif	
-	PPU.VMA.Address += PPU.VMA.Increment;
-    }
-//    Memory.FillRAM [0x2118] = Byte;
-}
-
-STATIC inline void REGISTER_2118_tile (uint8 Byte)
-{
-    uint32 address;
-    uint32 rem = PPU.VMA.Address & PPU.VMA.Mask1;
-    address = (((PPU.VMA.Address & ~PPU.VMA.Mask1) +
-		 (rem >> PPU.VMA.Shift) +
-		 ((rem & (PPU.VMA.FullGraphicCount - 1)) << 3)) << 1) & 0xffff;
-    Memory.VRAM [address] = Byte;
-    IPPU.TileCached [TILE_2BIT][address >> 4] = FALSE;
-    IPPU.TileCached [TILE_4BIT][address >> 5] = FALSE;
-    IPPU.TileCached [TILE_8BIT][address >> 6] = FALSE;
-    if (!PPU.VMA.High)
-	PPU.VMA.Address += PPU.VMA.Increment;
-//    Memory.FillRAM [0x2118] = Byte;
-}
-
-STATIC inline void REGISTER_2118_linear (uint8 Byte)
-{
-    uint32 address;
-    Memory.VRAM[address = (PPU.VMA.Address << 1) & 0xFFFF] = Byte;
-    IPPU.TileCached [TILE_2BIT][address >> 4] = FALSE;
-    IPPU.TileCached [TILE_4BIT][address >> 5] = FALSE;
-    IPPU.TileCached [TILE_8BIT][address >> 6] = FALSE;
-    if (!PPU.VMA.High)
-	PPU.VMA.Address += PPU.VMA.Increment;
-//    Memory.FillRAM [0x2118] = Byte;
-}
-
-STATIC inline void REGISTER_2119 (uint8 Byte)
-{
-    uint32 address;
-    if (PPU.VMA.FullGraphicCount)
-    {
-	uint32 rem = PPU.VMA.Address & PPU.VMA.Mask1;
-	address = ((((PPU.VMA.Address & ~PPU.VMA.Mask1) +
-		    (rem >> PPU.VMA.Shift) +
-		    ((rem & (PPU.VMA.FullGraphicCount - 1)) << 3)) << 1) + 1) & 0xFFFF;
-	Memory.VRAM [address] = Byte;
-    }
-    else
-    {
-	Memory.VRAM[address = ((PPU.VMA.Address << 1) + 1) & 0xFFFF] = Byte;
-    }
-    IPPU.TileCached [TILE_2BIT][address >> 4] = FALSE;
-    IPPU.TileCached [TILE_4BIT][address >> 5] = FALSE;
-    IPPU.TileCached [TILE_8BIT][address >> 6] = FALSE;
-    if (PPU.VMA.High)
-    {
-#ifdef DEBUGGER
-	if (Settings.TraceVRAM && !CPU.InDMA)
-	{
-	    printf ("VRAM write word: $%04X (%d,%d)\n", PPU.VMA.Address,
-		    Memory.FillRAM[0x2115] & 3,
-		    (Memory.FillRAM [0x2115] & 0x0c) >> 2);
-	}
-#endif	
-	PPU.VMA.Address += PPU.VMA.Increment;
-    }
-//    Memory.FillRAM [0x2119] = Byte;
-}
-
-STATIC inline void REGISTER_2119_tile (uint8 Byte)
-{
-    uint32 rem = PPU.VMA.Address & PPU.VMA.Mask1;
-    uint32 address = ((((PPU.VMA.Address & ~PPU.VMA.Mask1) +
-		    (rem >> PPU.VMA.Shift) +
-		    ((rem & (PPU.VMA.FullGraphicCount - 1)) << 3)) << 1) + 1) & 0xFFFF;
-    Memory.VRAM [address] = Byte;
-    IPPU.TileCached [TILE_2BIT][address >> 4] = FALSE;
-    IPPU.TileCached [TILE_4BIT][address >> 5] = FALSE;
-    IPPU.TileCached [TILE_8BIT][address >> 6] = FALSE;
-    if (PPU.VMA.High)
-	PPU.VMA.Address += PPU.VMA.Increment;
-//    Memory.FillRAM [0x2119] = Byte;
-}
-
-STATIC inline void REGISTER_2119_linear (uint8 Byte)
-{
-    uint32 address;
-    Memory.VRAM[address = ((PPU.VMA.Address << 1) + 1) & 0xFFFF] = Byte;
-    IPPU.TileCached [TILE_2BIT][address >> 4] = FALSE;
-    IPPU.TileCached [TILE_4BIT][address >> 5] = FALSE;
-    IPPU.TileCached [TILE_8BIT][address >> 6] = FALSE;
-    if (PPU.VMA.High)
-	PPU.VMA.Address += PPU.VMA.Increment;
-//    Memory.FillRAM [0x2119] = Byte;
-}
-
-STATIC inline void REGISTER_2122(uint8 Byte)
-{
-    // CG-RAM (palette) write
-
-    if (PPU.CGFLIP)
-    {
-	if ((Byte & 0x7f) != (PPU.CGDATA[PPU.CGADD] >> 8))
-	{
-	    if (Settings.SixteenBit)
-		FLUSH_REDRAW ();
-	    PPU.CGDATA[PPU.CGADD] &= 0x00FF;
-	    PPU.CGDATA[PPU.CGADD] |= (Byte & 0x7f) << 8;
-	    IPPU.ColorsChanged = TRUE;
-	    if (Settings.SixteenBit)
-	    {
-		IPPU.Blue [PPU.CGADD] = IPPU.XB [(Byte >> 2) & 0x1f];
-		IPPU.Green [PPU.CGADD] = IPPU.XB [(PPU.CGDATA[PPU.CGADD] >> 5) & 0x1f];
-		IPPU.ScreenColors [PPU.CGADD] = (uint16) BUILD_PIXEL (IPPU.Red [PPU.CGADD],
-							     IPPU.Green [PPU.CGADD],
-							     IPPU.Blue [PPU.CGADD]);
-	    }
-	}
-	PPU.CGADD++;
-    }
-    else
-    {
-	if (Byte != (uint8) (PPU.CGDATA[PPU.CGADD] & 0xff))
-	{
-	    if (Settings.SixteenBit)
-		FLUSH_REDRAW ();
-	    PPU.CGDATA[PPU.CGADD] &= 0x7F00;
-	    PPU.CGDATA[PPU.CGADD] |= Byte;
-	    IPPU.ColorsChanged = TRUE;
-	    if (Settings.SixteenBit)
-	    {
-		IPPU.Red [PPU.CGADD] = IPPU.XB [Byte & 0x1f];
-		IPPU.Green [PPU.CGADD] = IPPU.XB [(PPU.CGDATA[PPU.CGADD] >> 5) & 0x1f];
-		IPPU.ScreenColors [PPU.CGADD] = (uint16) BUILD_PIXEL (IPPU.Red [PPU.CGADD],
-							     IPPU.Green [PPU.CGADD],
-							     IPPU.Blue [PPU.CGADD]);
-	    }
-	}
-    }
-    PPU.CGFLIP ^= 1;
-//    Memory.FillRAM [0x2122] = Byte;
-}
-
-STATIC inline void REGISTER_2180(uint8 Byte)
-{
-    Memory.RAM[PPU.WRAM++] = Byte;
-    PPU.WRAM &= 0x1FFFF;
-    Memory.FillRAM [0x2180] = Byte;
-}
-
+extern uint8 REGISTER_4212();
+extern void FLUSH_REDRAW ();
+extern void REGISTER_2104 (uint8 byte);
+extern void REGISTER_2118 (uint8 Byte);
+extern void REGISTER_2118_tile (uint8 Byte);
+extern void REGISTER_2118_linear (uint8 Byte);
+extern void REGISTER_2119 (uint8 Byte);
+extern void REGISTER_2119_tile (uint8 Byte);
+extern void REGISTER_2119_linear (uint8 Byte);
+extern void REGISTER_2122(uint8 Byte);
+extern void REGISTER_2180(uint8 Byte);
 
 //Platform specific input functions used by PPU.CPP
 void JustifierButtons(uint32&);
diff --git a/source/sa1cpu.cpp b/source/sa1cpu.cpp
index 1532f57..655fa5c 100644
--- a/source/sa1cpu.cpp
+++ b/source/sa1cpu.cpp
@@ -134,6 +134,69 @@
 #define StackRelative SA1StackRelative
 #define StackRelativeIndirectIndexed SA1StackRelativeIndirectIndexed
 
+#define SetZN16 SA1SetZN16
+#define SetZN8 SA1SetZN8
+#define ADC8 SA1ADC8
+#define ADC16 SA1ADC16
+#define AND16 SA1AND16
+#define AND8 SA1AND8
+#define A_ASL16 SA1A_ASL16
+#define A_ASL8 SA1A_ASL8
+#define ASL16 SA1ASL16
+#define ASL8 SA1ASL8
+#define BIT16 SA1BIT16
+#define BIT8 SA1BIT8
+#define CMP16 SA1CMP16
+#define CMP8 SA1CMP8
+#define CMX16 SA1CMX16
+#define CMX8 SA1CMX8
+#define CMY16 SA1CMY16
+#define CMY8 SA1CMY8
+#define A_DEC16 SA1A_DEC16
+#define A_DEC8 SA1A_DEC8
+#define DEC16 SA1DEC16
+#define DEC8 SA1DEC8
+#define EOR16 SA1EOR16
+#define EOR8 SA1EOR8
+#define A_INC16 SA1A_INC16
+#define A_INC8 SA1A_INC8
+#define INC16 SA1INC16
+#define INC8 SA1INC8
+#define LDA16 SA1LDA16
+#define LDA8 SA1LDA8
+#define LDX16 SA1LDX16
+#define LDX8 SA1LDX8
+#define LDY16 SA1LDY16
+#define LDY8 SA1LDY8
+#define A_LSR16 SA1A_LSR16
+#define A_LSR8 SA1A_LSR8
+#define LSR16 SA1LSR16
+#define LSR8 SA1LSR8
+#define ORA16 SA1ORA16
+#define ORA8 SA1ORA8
+#define A_ROL16 SA1A_ROL16
+#define A_ROL8 SA1A_ROL8
+#define ROL16 SA1ROL16
+#define ROL8 SA1ROL8
+#define A_ROR16 SA1A_ROR16
+#define A_ROR8 SA1A_ROR8
+#define ROR16 SA1ROR16
+#define ROR8 SA1ROR8
+#define SBC16 SA1SBC16
+#define SBC8 SA1SBC8
+#define STA16 SA1STA16
+#define STA8 SA1STA8
+#define STX16 SA1STX16
+#define STX8 SA1STX8
+#define STY16 SA1STY16
+#define STY8 SA1STY8
+#define STZ16 SA1STZ16
+#define STZ8 SA1STZ8
+#define TSB16 SA1TSB16
+#define TSB8 SA1TSB8
+#define TRB16 SA1TRB16
+#define TRB8 SA1TRB8
+
 //#undef CPU_SHUTDOWN
 #undef VAR_CYCLES
 #define SA1_OPCODES
diff --git a/source/tile.cpp b/source/tile.cpp
index 27e61f0..d4cdcbf 100644
--- a/source/tile.cpp
+++ b/source/tile.cpp
@@ -216,7 +216,7 @@ uint8 ConvertTile (uint8 *pCache, uint32 TileAddr)
     return (non_zero ? TRUE : BLANK_TILE);
 }
 
-inline void WRITE_4PIXELS (uint32 Offset, uint8 *Pixels)
+void WRITE_4PIXELS (uint32 Offset, uint8 *Pixels)
 {
     uint8 Pixel;
     uint8 *Screen = GFX.S + Offset;
@@ -236,7 +236,7 @@ inline void WRITE_4PIXELS (uint32 Offset, uint8 *Pixels)
 #undef FN
 }
 
-inline void WRITE_4PIXELS_FLIPPED (uint32 Offset, uint8 *Pixels)
+void WRITE_4PIXELS_FLIPPED (uint32 Offset, uint8 *Pixels)
 {
     uint8 Pixel;
     uint8 *Screen = GFX.S + Offset;
@@ -256,7 +256,7 @@ inline void WRITE_4PIXELS_FLIPPED (uint32 Offset, uint8 *Pixels)
 #undef FN
 }
 
-inline void WRITE_4PIXELSx2 (uint32 Offset, uint8 *Pixels)
+void WRITE_4PIXELSx2 (uint32 Offset, uint8 *Pixels)
 {
     uint8 Pixel;
     uint8 *Screen = GFX.S + Offset;
@@ -276,7 +276,7 @@ inline void WRITE_4PIXELSx2 (uint32 Offset, uint8 *Pixels)
 #undef FN
 }
 
-inline void WRITE_4PIXELS_FLIPPEDx2 (uint32 Offset, uint8 *Pixels)
+void WRITE_4PIXELS_FLIPPEDx2 (uint32 Offset, uint8 *Pixels)
 {
     uint8 Pixel;
     uint8 *Screen = GFX.S + Offset;
@@ -296,7 +296,7 @@ inline void WRITE_4PIXELS_FLIPPEDx2 (uint32 Offset, uint8 *Pixels)
 #undef FN
 }
 
-inline void WRITE_4PIXELSx2x2 (uint32 Offset, uint8 *Pixels)
+void WRITE_4PIXELSx2x2 (uint32 Offset, uint8 *Pixels)
 {
     uint8 Pixel;
     uint8 *Screen = GFX.S + Offset;
@@ -318,7 +318,7 @@ inline void WRITE_4PIXELSx2x2 (uint32 Offset, uint8 *Pixels)
 #undef FN
 }
 
-inline void WRITE_4PIXELS_FLIPPEDx2x2 (uint32 Offset, uint8 *Pixels)
+void WRITE_4PIXELS_FLIPPEDx2x2 (uint32 Offset, uint8 *Pixels)
 {
     uint8 Pixel;
     uint8 *Screen = GFX.S + Offset;
@@ -417,7 +417,7 @@ void DrawLargePixel (uint32 Tile, uint32 Offset,
     RENDER_TILE_LARGE (((uint8) GFX.ScreenColors [pixel]), PLOT_PIXEL)
 }
 
-inline void WRITE_4PIXELS16 (uint32 Offset, uint8 *Pixels)
+void WRITE_4PIXELS16 (uint32 Offset, uint8 *Pixels)
 {
     uint32 Pixel;
     uint16 *Screen = (uint16 *) GFX.S + Offset;
@@ -437,7 +437,7 @@ inline void WRITE_4PIXELS16 (uint32 Offset, uint8 *Pixels)
 #undef FN
 }
 
-inline void WRITE_4PIXELS16_FLIPPED (uint32 Offset, uint8 *Pixels)
+void WRITE_4PIXELS16_FLIPPED (uint32 Offset, uint8 *Pixels)
 {
     uint32 Pixel;
     uint16 *Screen = (uint16 *) GFX.S + Offset;
@@ -457,7 +457,7 @@ inline void WRITE_4PIXELS16_FLIPPED (uint32 Offset, uint8 *Pixels)
 #undef FN
 }
 
-inline void WRITE_4PIXELS16x2 (uint32 Offset, uint8 *Pixels)
+void WRITE_4PIXELS16x2 (uint32 Offset, uint8 *Pixels)
 {
     uint32 Pixel;
     uint16 *Screen = (uint16 *) GFX.S + Offset;
@@ -477,7 +477,7 @@ inline void WRITE_4PIXELS16x2 (uint32 Offset, uint8 *Pixels)
 #undef FN
 }
 
-inline void WRITE_4PIXELS16_FLIPPEDx2 (uint32 Offset, uint8 *Pixels)
+void WRITE_4PIXELS16_FLIPPEDx2 (uint32 Offset, uint8 *Pixels)
 {
     uint32 Pixel;
     uint16 *Screen = (uint16 *) GFX.S + Offset;
@@ -497,7 +497,7 @@ inline void WRITE_4PIXELS16_FLIPPEDx2 (uint32 Offset, uint8 *Pixels)
 #undef FN
 }
 
-inline void WRITE_4PIXELS16x2x2 (uint32 Offset, uint8 *Pixels)
+void WRITE_4PIXELS16x2x2 (uint32 Offset, uint8 *Pixels)
 {
     uint32 Pixel;
     uint16 *Screen = (uint16 *) GFX.S + Offset;
@@ -519,7 +519,7 @@ inline void WRITE_4PIXELS16x2x2 (uint32 Offset, uint8 *Pixels)
 #undef FN
 }
 
-inline void WRITE_4PIXELS16_FLIPPEDx2x2 (uint32 Offset, uint8 *Pixels)
+void WRITE_4PIXELS16_FLIPPEDx2x2 (uint32 Offset, uint8 *Pixels)
 {
     uint32 Pixel;
     uint16 *Screen = (uint16 *) GFX.S + Offset;
@@ -614,7 +614,7 @@ void DrawLargePixel16 (uint32 Tile, uint32 Offset,
     RENDER_TILE_LARGE (GFX.ScreenColors [pixel], PLOT_PIXEL)
 }
 
-inline void WRITE_4PIXELS16_ADD (uint32 Offset, uint8 *Pixels)
+void WRITE_4PIXELS16_ADD (uint32 Offset, uint8 *Pixels)
 {
     uint32 Pixel;
     uint16 *Screen = (uint16 *) GFX.S + Offset;
@@ -646,7 +646,7 @@ inline void WRITE_4PIXELS16_ADD (uint32 Offset, uint8 *Pixels)
 #undef FN
 }
 
-inline void WRITE_4PIXELS16_FLIPPED_ADD (uint32 Offset, uint8 *Pixels)
+void WRITE_4PIXELS16_FLIPPED_ADD (uint32 Offset, uint8 *Pixels)
 {
     uint32 Pixel;
     uint16 *Screen = (uint16 *) GFX.S + Offset;
@@ -678,7 +678,7 @@ inline void WRITE_4PIXELS16_FLIPPED_ADD (uint32 Offset, uint8 *Pixels)
 #undef FN
 }
 
-inline void WRITE_4PIXELS16_ADD1_2 (uint32 Offset, uint8 *Pixels)
+void WRITE_4PIXELS16_ADD1_2 (uint32 Offset, uint8 *Pixels)
 {
     uint32 Pixel;
     uint16 *Screen = (uint16 *) GFX.S + Offset;
@@ -710,7 +710,7 @@ inline void WRITE_4PIXELS16_ADD1_2 (uint32 Offset, uint8 *Pixels)
 #undef FN
 }
 
-inline void WRITE_4PIXELS16_FLIPPED_ADD1_2 (uint32 Offset, uint8 *Pixels)
+void WRITE_4PIXELS16_FLIPPED_ADD1_2 (uint32 Offset, uint8 *Pixels)
 {
     uint32 Pixel;
     uint16 *Screen = (uint16 *) GFX.S + Offset;
@@ -742,7 +742,7 @@ inline void WRITE_4PIXELS16_FLIPPED_ADD1_2 (uint32 Offset, uint8 *Pixels)
 #undef FN
 }
 
-inline void WRITE_4PIXELS16_SUB (uint32 Offset, uint8 *Pixels)
+void WRITE_4PIXELS16_SUB (uint32 Offset, uint8 *Pixels)
 {
     uint32 Pixel;
     uint16 *Screen = (uint16 *) GFX.S + Offset;
@@ -774,7 +774,7 @@ inline void WRITE_4PIXELS16_SUB (uint32 Offset, uint8 *Pixels)
 #undef FN
 }
 
-inline void WRITE_4PIXELS16_FLIPPED_SUB (uint32 Offset, uint8 *Pixels)
+void WRITE_4PIXELS16_FLIPPED_SUB (uint32 Offset, uint8 *Pixels)
 {
     uint32 Pixel;
     uint16 *Screen = (uint16 *) GFX.S + Offset;
@@ -806,7 +806,7 @@ inline void WRITE_4PIXELS16_FLIPPED_SUB (uint32 Offset, uint8 *Pixels)
 #undef FN
 }
 
-inline void WRITE_4PIXELS16_SUB1_2 (uint32 Offset, uint8 *Pixels)
+void WRITE_4PIXELS16_SUB1_2 (uint32 Offset, uint8 *Pixels)
 {
     uint32 Pixel;
     uint16 *Screen = (uint16 *) GFX.S + Offset;
@@ -838,7 +838,7 @@ inline void WRITE_4PIXELS16_SUB1_2 (uint32 Offset, uint8 *Pixels)
 #undef FN
 }
 
-inline void WRITE_4PIXELS16_FLIPPED_SUB1_2 (uint32 Offset, uint8 *Pixels)
+void WRITE_4PIXELS16_FLIPPED_SUB1_2 (uint32 Offset, uint8 *Pixels)
 {
     uint32 Pixel;
     uint16 *Screen = (uint16 *) GFX.S + Offset;
@@ -951,7 +951,7 @@ void DrawClippedTile16Sub1_2 (uint32 Tile, uint32 Offset,
     RENDER_CLIPPED_TILE(WRITE_4PIXELS16_SUB1_2, WRITE_4PIXELS16_FLIPPED_SUB1_2, 4)
 }
 
-inline void WRITE_4PIXELS16_ADDF1_2 (uint32 Offset, uint8 *Pixels)
+void WRITE_4PIXELS16_ADDF1_2 (uint32 Offset, uint8 *Pixels)
 {
     uint32 Pixel;
     uint16 *Screen = (uint16 *) GFX.S + Offset;
@@ -977,7 +977,7 @@ inline void WRITE_4PIXELS16_ADDF1_2 (uint32 Offset, uint8 *Pixels)
 #undef FN
 }
 
-inline void WRITE_4PIXELS16_FLIPPED_ADDF1_2 (uint32 Offset, uint8 *Pixels)
+void WRITE_4PIXELS16_FLIPPED_ADDF1_2 (uint32 Offset, uint8 *Pixels)
 {
     uint32 Pixel;
     uint16 *Screen = (uint16 *) GFX.S + Offset;
@@ -1003,7 +1003,7 @@ inline void WRITE_4PIXELS16_FLIPPED_ADDF1_2 (uint32 Offset, uint8 *Pixels)
 #undef FN
 }
 
-inline void WRITE_4PIXELS16_SUBF1_2 (uint32 Offset, uint8 *Pixels)
+void WRITE_4PIXELS16_SUBF1_2 (uint32 Offset, uint8 *Pixels)
 {
     uint32 Pixel;
     uint16 *Screen = (uint16 *) GFX.S + Offset;
@@ -1029,7 +1029,7 @@ inline void WRITE_4PIXELS16_SUBF1_2 (uint32 Offset, uint8 *Pixels)
 #undef FN
 }
 
-inline void WRITE_4PIXELS16_FLIPPED_SUBF1_2 (uint32 Offset, uint8 *Pixels)
+void WRITE_4PIXELS16_FLIPPED_SUBF1_2 (uint32 Offset, uint8 *Pixels)
 {
     uint32 Pixel;
     uint16 *Screen = (uint16 *) GFX.S + Offset;