Merge pull request libretro#92 from jSTE0/microopt

Author: inactive123

source/dsp1.c

@@ -352,7 +352,7 @@ void DSP1SetByte(uint8_t byte, uint16_t address)
                   Op11m = (int16_t)(DSP1.parameters [0] | (DSP1.parameters[1] << 8));
                   Op11Zr = (int16_t)(DSP1.parameters [2] | (DSP1.parameters[3] << 8));
                   Op11Yr = (int16_t)(DSP1.parameters [4] | (DSP1.parameters[5] << 8));
-                  Op11Xr = (int16_t)(DSP1.parameters [7] | (DSP1.parameters[7] << 8));
+                  Op11Xr = (int16_t)(DSP1.parameters [6] | (DSP1.parameters[7] << 8));
                   DSPOp11();
                   break;
                case 0x25:

source/dsp1emu.c

@@ -187,11 +187,19 @@ void DSP1_Inverse(int16_t Coefficient, int16_t Exponent, int16_t* iCoefficient,
       }
 
       /* 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)
@@ -336,9 +344,18 @@ int16_t DSP1_Cos(int16_t Angle)
 
 void DSP1_Normalize(int16_t m, int16_t* Coefficient, int16_t* Exponent)
 {
-   int16_t i = 0x4000;
    int16_t e = 0;
 
+#ifdef __GNUC__
+   int16_t n = m < 0 ? ~m : m;
+
+   if (n == 0)
+      e = 15;
+   else
+      e = __builtin_clz(n) - (8 * sizeof(int) - 15);
+#else
+   int16_t i = 0x4000;
+
    if (m < 0)
    {
       while ((m & i) && i)
@@ -355,6 +372,7 @@ void DSP1_Normalize(int16_t m, int16_t* Coefficient, int16_t* Exponent)
          e++;
       }
    }
+#endif
 
    if (e > 0)
       *Coefficient = m * DSP1ROM[0x21 + e] << 1;
@@ -368,9 +386,18 @@ void DSP1_NormalizeDouble(int32_t Product, int16_t* Coefficient, int16_t* Expone
 {
    int16_t n = Product & 0x7fff;
    int16_t m = Product >> 15;
-   int16_t i = 0x4000;
    int16_t e = 0;
 
+#ifdef __GNUC__
+   int16_t t = m < 0 ? ~m : m;
+
+   if (t == 0)
+      e = 15;
+   else
+      e = __builtin_clz(t) - (8 * sizeof(int) - 15);
+#else
+   int16_t i = 0x4000;
+
    if (m < 0)
    {
       while ((m & i) && i)
@@ -387,6 +414,7 @@ void DSP1_NormalizeDouble(int32_t Product, int16_t* Coefficient, int16_t* Expone
          e++;
       }
    }
+#endif
 
    if (e > 0)
    {
@@ -396,6 +424,14 @@ void DSP1_NormalizeDouble(int32_t Product, int16_t* Coefficient, int16_t* Expone
          *Coefficient += n * DSP1ROM[0x0040 - e] >> 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)
@@ -414,6 +450,7 @@ void DSP1_NormalizeDouble(int32_t Product, int16_t* Coefficient, int16_t* Expone
                e++;
             }
          }
+#endif
 
          if (e > 15)
             *Coefficient = n * DSP1ROM[0x0012 + e] << 1;

source/fxemu.c

@@ -2,6 +2,7 @@
 
 #include "fxemu.h"
 #include "fxinst.h"
+#include "memmap.h"
 #include <stdlib.h>
 #include <string.h>
 
@@ -30,33 +31,27 @@ void fx_updateRamBank(uint8_t Byte)
 
 static INLINE void fx_readRegisterSpaceForCheck(void)
 {
-   R15 = GSU.pvRegisters[30];
-   R15 |= ((uint32_t) GSU.pvRegisters[31]) << 8;
+   R15 = (uint32_t) READ_WORD(&GSU.pvRegisters[30]);
 }
 
 static void fx_readRegisterSpaceForUse(void)
 {
-   static uint32_t avHeight[] = { 128, 160, 192, 256 };
-   static uint32_t avMult[] = { 16, 32, 32, 64 };
+   static const uint32_t avHeight[] = { 128, 160, 192, 256 };
+   static const uint32_t avMult[] = { 16, 32, 32, 64 };
    int32_t i;
    uint8_t* p = GSU.pvRegisters;
 
    /* Update R0 - R14 */
-   for (i = 0; i < 15; i++)
-   {
-      GSU.avReg[i] = *p++;
-      GSU.avReg[i] += ((uint32_t)(*p++)) << 8;
-   }
+   for (i = 0; i < 15; i++, p += 2)
+      GSU.avReg[i] = (uint32_t) READ_WORD(p);
 
    /* 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.vStatusReg = (uint32_t) READ_WORD(&GSU.pvRegisters[GSU_SFR]);
    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];
-   GSU.vCacheBaseReg |= ((uint32_t)p[GSU_CBR + 1]) << 8;
+   GSU.vCacheBaseReg = (uint32_t) READ_WORD(&p[GSU_CBR]);
 
    /* Update status register variables */
    GSU.vZero = !(GSU.vStatusReg & FLG_Z);
@@ -146,19 +141,15 @@ void fx_computeScreenPointers(void)
 
 static INLINE void fx_writeRegisterSpaceAfterCheck(void)
 {
-   GSU.pvRegisters[30] = (uint8_t) R15;
-   GSU.pvRegisters[31] = (uint8_t) (R15 >> 8);
+   WRITE_WORD(&GSU.pvRegisters[30], R15);
 }
 
 static void fx_writeRegisterSpaceAfterUse(void)
 {
    int32_t i;
    uint8_t* p = GSU.pvRegisters;
-   for (i = 0; i < 15; i++)
-   {
-      *p++ = (uint8_t)GSU.avReg[i];
-      *p++ = (uint8_t)(GSU.avReg[i] >> 8);
-   }
+   for (i = 0; i < 15; i++, p += 2)
+      WRITE_WORD(p, GSU.avReg[i]);
 
    /* Update status register */
    if (USEX16(GSU.vZero) == 0)
@@ -179,13 +170,11 @@ 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);
+   WRITE_WORD(&p[GSU_SFR], GSU.vStatusReg);
    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;
-   p[GSU_CBR + 1] = (uint8_t)(GSU.vCacheBaseReg >> 8);
+   WRITE_WORD(&p[GSU_CBR], GSU.vCacheBaseReg);
 }
 
 /* Reset the FxChip */

source/gfx.c

@@ -423,7 +423,7 @@ void S9xStartScreenRefresh(void)
       GFX.Delta = (GFX.SubScreen - GFX.Screen) >> 1;
    }
 
-   if (++IPPU.FrameCount % Memory.ROMFramesPerSecond == 0)
+   if (++IPPU.FrameCount == (uint32_t)Memory.ROMFramesPerSecond)
       IPPU.FrameCount = 0;
 }