Merge pull request #48 from jSTE0/microopt

Microopt

src/dsp1.c

@@ -588,7 +588,7 @@ void DSP1SetByte(uint8 byte, uint16 address)
                   Op11m = (int16)(DSP1.parameters [0] | (DSP1.parameters[1] << 8));
                   Op11Zr = (int16)(DSP1.parameters [2] | (DSP1.parameters[3] << 8));
                   Op11Yr = (int16)(DSP1.parameters [4] | (DSP1.parameters[5] << 8));
-                  Op11Xr = (int16)(DSP1.parameters [7] | (DSP1.parameters[7] << 8));
+                  Op11Xr = (int16)(DSP1.parameters [6] | (DSP1.parameters[7] << 8));
 
                   DSPOp11();
                   break;

src/dsp1emu.c

@@ -255,11 +255,19 @@ void DSP1_Inverse(short Coefficient, short Exponent, short* iCoefficient, short*
       }
 
       // Step Three: Normalize
+#ifdef __GNUC__
+      {
+          const int shift = __builtin_clz(Coefficient) - (8 * sizeof(int) - 15);
+          Coefficient <<= shift;
+          Exponent -= shift;
+      }
+#else
       while (Coefficient < 0x4000)
       {
          Coefficient <<= 1;
          Exponent--;
       }
+#endif
 
       // Step Four: Special Case
       if (Coefficient == 0x4000)
@@ -395,9 +403,18 @@ short DSP1_Cos(short Angle)
 
 void DSP1_Normalize(short m, short* Coefficient, short* Exponent)
 {
-   short i = 0x4000;
    short e = 0;
 
+#ifdef __GNUC__
+   int n = m < 0 ? ~m : m;
+
+   if (n == 0)
+      e = 15;
+   else
+      e = __builtin_clz(n) - (8 * sizeof(int) - 15);
+#else
+   short i = 0x4000;
+
    if (m < 0)
       while ((m & i) && i)
       {
@@ -410,6 +427,7 @@ void DSP1_Normalize(short m, short* Coefficient, short* Exponent)
          i >>= 1;
          e++;
       }
+#endif
 
    if (e > 0)
       *Coefficient = m * DSP1ROM[(0x21 + e) & 1023] << 1;
@@ -423,9 +441,18 @@ void DSP1_Normalizefloat(int Product, short* Coefficient, short* Exponent)
 {
    short n = Product & 0x7fff;
    short m = Product >> 15;
-   short i = 0x4000;
    short e = 0;
 
+#ifdef __GNUC__
+   short t = m < 0 ? ~m : m;
+
+   if (t == 0)
+      e = 15;
+   else
+      e = __builtin_clz(t) - (8 * sizeof(int) - 15);
+#else
+   short i = 0x4000;
+
    if (m < 0)
       while ((m & i) && i)
       {
@@ -438,6 +465,7 @@ void DSP1_Normalizefloat(int Product, short* Coefficient, short* Exponent)
          i >>= 1;
          e++;
       }
+#endif
 
    if (e > 0)
    {
@@ -447,6 +475,14 @@ void DSP1_Normalizefloat(int Product, short* Coefficient, short* Exponent)
          *Coefficient += n * DSP1ROM[(0x0040 - e) & 1023] >> 15;
       else
       {
+#ifdef __GNUC__
+         t = m < 0 ? ~(n | 0x8000) : n;
+
+         if (t == 0)
+            e += 15;
+         else
+            e += __builtin_clz(t) - (8 * sizeof(int) - 15);
+#else
          i = 0x4000;
 
          if (m < 0)
@@ -461,6 +497,7 @@ void DSP1_Normalizefloat(int Product, short* Coefficient, short* Exponent)
                i >>= 1;
                e++;
             }
+#endif
 
          if (e > 15)
             *Coefficient = n * DSP1ROM[(0x0012 + e) & 1023] << 1;

src/fxemu.c

@@ -90,6 +90,7 @@
 *******************************************************************************/
 #include "fxemu.h"
 #include "fxinst.h"
+#include "memmap.h"
 #include <stdlib.h>
 #include <string.h>
 #include <stdio.h>
@@ -245,28 +246,23 @@ static INLINE void fx_readRegisterSpace()
 {
    int i;
    uint8* p;
-   static uint32 avHeight[] = { 128, 160, 192, 256 };
-   static uint32 avMult[] = { 16, 32, 32, 64 };
+   static const uint32 avHeight[] = { 128, 160, 192, 256 };
+   static const uint32 avMult[] = { 16, 32, 32, 64 };
 
    GSU.vErrorCode = 0;
 
    /* Update R0-R15 */
    p = GSU.pvRegisters;
-   for (i = 0; i < 16; i++)
-   {
-      GSU.avReg[i] = *p++;
-      GSU.avReg[i] += ((uint32)(*p++)) << 8;
-   }
+   for (i = 0; i < 16; i++, p += 2)
+      GSU.avReg[i] = (uint32)READ_WORD(p);
 
    /* Update other registers */
    p = GSU.pvRegisters;
-   GSU.vStatusReg = (uint32)p[GSU_SFR];
-   GSU.vStatusReg |= ((uint32)p[GSU_SFR + 1]) << 8;
+   GSU.vStatusReg = (uint32)READ_WORD(&p[GSU_SFR]);
    GSU.vPrgBankReg = (uint32)p[GSU_PBR];
    GSU.vRomBankReg = (uint32)p[GSU_ROMBR];
    GSU.vRamBankReg = ((uint32)p[GSU_RAMBR]) & (FX_RAM_BANKS - 1);
-   GSU.vCacheBaseReg = (uint32)p[GSU_CBR];
-   GSU.vCacheBaseReg |= ((uint32)p[GSU_CBR + 1]) << 8;
+   GSU.vCacheBaseReg = (uint32)READ_WORD(&p[GSU_CBR]);
 
    /* Update status register variables */
    GSU.vZero = !(GSU.vStatusReg & FLG_Z);
@@ -459,11 +455,8 @@ static INLINE void fx_writeRegisterSpace()
    uint8* p;
 
    p = GSU.pvRegisters;
-   for (i = 0; i < 16; i++)
-   {
-      *p++ = (uint8)GSU.avReg[i];
-      *p++ = (uint8)(GSU.avReg[i] >> 8);
-   }
+   for (i = 0; i < 16; i++, p += 2)
+      WRITE_WORD(p, GSU.avReg[i]);
 
    /* Update status register */
    if (USEX16(GSU.vZero) == 0) SF(Z);
@@ -476,13 +469,11 @@ static INLINE void fx_writeRegisterSpace()
    else CF(CY);
 
    p = GSU.pvRegisters;
-   p[GSU_SFR] = (uint8)GSU.vStatusReg;
-   p[GSU_SFR + 1] = (uint8)(GSU.vStatusReg >> 8);
+   WRITE_WORD(&p[GSU_SFR], GSU.vStatusReg);
    p[GSU_PBR] = (uint8)GSU.vPrgBankReg;
    p[GSU_ROMBR] = (uint8)GSU.vRomBankReg;
    p[GSU_RAMBR] = (uint8)GSU.vRamBankReg;
-   p[GSU_CBR] = (uint8)GSU.vCacheBaseReg;
-   p[GSU_CBR + 1] = (uint8)(GSU.vCacheBaseReg >> 8);
+   WRITE_WORD(&p[GSU_CBR], GSU.vCacheBaseReg);
 
    fx_restoreCache();
 }

src/gfx.c

@@ -579,7 +579,7 @@ void S9xStartScreenRefresh()
       GFX.DepthDelta = GFX.SubZBuffer - GFX.ZBuffer;
       GFX.Delta = (GFX.SubScreen - GFX.Screen) >> 1;
    }
-   if (++IPPU.FrameCount % Memory.ROMFramesPerSecond == 0)
+   if (++IPPU.FrameCount == (uint32)Memory.ROMFramesPerSecond)
    {
       IPPU.DisplayedRenderedFrameCount = IPPU.RenderedFramesCount;
       IPPU.RenderedFramesCount = 0;

src/gfx16.c

@@ -597,7 +597,7 @@ void S9xStartScreenRefresh()
       GFX.Delta = (GFX.SubScreen - GFX.Screen) >> 1;
    }
    
-   if (++IPPU.FrameCount % Memory.ROMFramesPerSecond == 0)
+   if (++IPPU.FrameCount == (uint32)Memory.ROMFramesPerSecond)
    {
       IPPU.DisplayedRenderedFrameCount = IPPU.RenderedFramesCount;
       IPPU.RenderedFramesCount = 0;