Use MSB_FIRST instead of LSB_FIRST

9 files changed, 88 insertions, 131 deletions

diff --git a/source/65c816.h b/source/65c816.h
index 505d185..406341a 100644
--- a/source/65c816.h
+++ b/source/65c816.h
@@ -146,15 +146,15 @@
 
 typedef union
 {
-#ifdef LSB_FIRST
+#ifdef MSB_FIRST
    struct
    {
-      uint8_t l, h;
+      uint8_t h, l;
    } B;
 #else
    struct
    {
-      uint8_t h, l;
+      uint8_t l, h;
    } B;
 #endif
    uint16_t W;
diff --git a/source/blargg_endian.h b/source/blargg_endian.h
index 8322fb1..177a081 100644
--- a/source/blargg_endian.h
+++ b/source/blargg_endian.h
@@ -65,75 +65,41 @@
    #define BLARGG_CPU_RISC 1
 #endif
 
-/* BLARGG_BIG_ENDIAN, BLARGG_LITTLE_ENDIAN: Determined automatically, otherwise only */
-/* one may be #defined to 1. Only needed if something actually depends on byte order. */
-#if !defined (BLARGG_BIG_ENDIAN) && !defined (BLARGG_LITTLE_ENDIAN)
-#ifdef __GLIBC__
-   /* GCC handles this for us */
-   #include <endian.h>
-   #if __BYTE_ORDER == __LITTLE_ENDIAN
-      #define BLARGG_LITTLE_ENDIAN 1
-   #elif __BYTE_ORDER == __BIG_ENDIAN
-      #define BLARGG_BIG_ENDIAN 1
-   #endif
-#else
-
-#if defined (LSB_FIRST) || defined (__LITTLE_ENDIAN__) || BLARGG_CPU_X86 ||  \
-                defined(ANDROID_X86) || defined(ANDROID_MIPS) || \
-   defined(__BLACKBERRY_QNX__) || (defined (LITTLE_ENDIAN) && LITTLE_ENDIAN+0 != 1234)
-   #define BLARGG_LITTLE_ENDIAN 1
-#endif
+#if BLARGG_NONPORTABLE
+   /* Optimized implementation if byte order is known */
 
-#if defined (MSB_FIRST)     || defined (__BIG_ENDIAN__) || defined (WORDS_BIGENDIAN) || \
-   defined (__sparc__)     ||  BLARGG_CPU_POWERPC || \
-   (defined (BIG_ENDIAN) && BIG_ENDIAN+0 != 4321)
-   #define BLARGG_BIG_ENDIAN 1
-#elif !defined(__mips__) || !defined(ANDROID_MIPS)
-   /* No endian specified; assume little-endian, since it's most common */
-   #define BLARGG_LITTLE_ENDIAN 1
-#endif
+#ifdef MSB_FIRST
+#if BLARGG_CPU_POWERPC
+   /* PowerPC has special byte-reversed instructions */
+#if defined (__SNC__)
+#define GET_LE16( addr )        (__builtin_lhbrx(addr, 0))
+#define GET_LE32( addr )        (__builtin_lwbrx(addr, 0))
+#define SET_LE16( addr, in )    (__builtin_sthbrx(in, addr, 0))
+#define SET_LE32( addr, in )    (__builtin_stwbrx(in, addr, 0))
+#elif defined (_XBOX360)
+#include <PPCIntrinsics.h>
+#define GET_LE16( addr )        (__loadshortbytereverse(0, addr))
+#define GET_LE32( addr )        (__loadwordbytereverse(0, addr))
+#define SET_LE16( addr, in )    (__storeshortbytereverse(in, 0, addr))
+#define SET_LE32( addr, in )    (__storewordbytereverse(in, 0, addr))
+#elif defined (__MWERKS__)
+#define GET_LE16( addr )        (__lhbrx( addr, 0 ))
+#define GET_LE32( addr )        (__lwbrx( addr, 0 ))
+#define SET_LE16( addr, in )    (__sthbrx( in, addr, 0 ))
+#define SET_LE32( addr, in )    (__stwbrx( in, addr, 0 ))
+#elif defined (__GNUC__)
+#define GET_LE16( addr )        ({unsigned ppc_lhbrx_; asm( "lhbrx %0,0,%1" : "=r" (ppc_lhbrx_) : "r" (addr), "0" (ppc_lhbrx_) ); ppc_lhbrx_;})
+#define GET_LE32( addr )        ({unsigned ppc_lwbrx_; asm( "lwbrx %0,0,%1" : "=r" (ppc_lwbrx_) : "r" (addr), "0" (ppc_lwbrx_) ); ppc_lwbrx_;})
+#define SET_LE16( addr, in )    ({asm( "sthbrx %0,0,%1" : : "r" (in), "r" (addr) );})
+#define SET_LE32( addr, in )    ({asm( "stwbrx %0,0,%1" : : "r" (in), "r" (addr) );})
 #endif
 #endif
-
-#if BLARGG_LITTLE_ENDIAN && BLARGG_BIG_ENDIAN
-   #undef BLARGG_LITTLE_ENDIAN
-   #undef BLARGG_BIG_ENDIAN
+#else
+#define GET_LE16( addr )        (*(uint16_t*) (addr))
+#define GET_LE32( addr )        (*(uint32_t*) (addr))
+#define SET_LE16( addr, data )  (void) (*(uint16_t*) (addr) = (data))
+#define SET_LE32( addr, data )  (void) (*(uint32_t*) (addr) = (data))
 #endif
-
-#if BLARGG_NONPORTABLE
-   /* Optimized implementation if byte order is known */
-   #if BLARGG_LITTLE_ENDIAN
-      #define GET_LE16( addr )        (*(uint16_t*) (addr))
-      #define GET_LE32( addr )        (*(uint32_t*) (addr))
-      #define SET_LE16( addr, data )  (void) (*(uint16_t*) (addr) = (data))
-      #define SET_LE32( addr, data )  (void) (*(uint32_t*) (addr) = (data))
-   #elif BLARGG_BIG_ENDIAN
-      #if BLARGG_CPU_POWERPC
-         /* PowerPC has special byte-reversed instructions */
-         #if defined (__SNC__)
-            #define GET_LE16( addr )        (__builtin_lhbrx(addr, 0))
-            #define GET_LE32( addr )        (__builtin_lwbrx(addr, 0))
-            #define SET_LE16( addr, in )    (__builtin_sthbrx(in, addr, 0))
-            #define SET_LE32( addr, in )    (__builtin_stwbrx(in, addr, 0))
-         #elif defined (_XBOX360)
-            #include <PPCIntrinsics.h>
-            #define GET_LE16( addr )        (__loadshortbytereverse(0, addr))
-            #define GET_LE32( addr )        (__loadwordbytereverse(0, addr))
-            #define SET_LE16( addr, in )    (__storeshortbytereverse(in, 0, addr))
-            #define SET_LE32( addr, in )    (__storewordbytereverse(in, 0, addr))
-         #elif defined (__MWERKS__)
-            #define GET_LE16( addr )        (__lhbrx( addr, 0 ))
-            #define GET_LE32( addr )        (__lwbrx( addr, 0 ))
-            #define SET_LE16( addr, in )    (__sthbrx( in, addr, 0 ))
-            #define SET_LE32( addr, in )    (__stwbrx( in, addr, 0 ))
-         #elif defined (__GNUC__)
-            #define GET_LE16( addr )        ({unsigned ppc_lhbrx_; asm( "lhbrx %0,0,%1" : "=r" (ppc_lhbrx_) : "r" (addr), "0" (ppc_lhbrx_) ); ppc_lhbrx_;})
-            #define GET_LE32( addr )        ({unsigned ppc_lwbrx_; asm( "lwbrx %0,0,%1" : "=r" (ppc_lwbrx_) : "r" (addr), "0" (ppc_lwbrx_) ); ppc_lwbrx_;})
-            #define SET_LE16( addr, in )    ({asm( "sthbrx %0,0,%1" : : "r" (in), "r" (addr) );})
-            #define SET_LE32( addr, in )    ({asm( "stwbrx %0,0,%1" : : "r" (in), "r" (addr) );})
-         #endif
-      #endif
-   #endif
 #else
 static INLINE unsigned get_le16( void const* p )
 {
diff --git a/source/gfx.c b/source/gfx.c
index 6c86ba5..2cd7584 100644
--- a/source/gfx.c
+++ b/source/gfx.c
@@ -268,24 +268,7 @@ bool S9xInitGFX()
          register uint32_t h = 0;
          register uint32_t l = 0;
 
-#if defined(LSB_FIRST)
-         if (i & 8)
-            h |= PixelOdd;
-         if (i & 4)
-            h |= PixelOdd << 8;
-         if (i & 2)
-            h |= PixelOdd << 16;
-         if (i & 1)
-            h |= PixelOdd << 24;
-         if (i & 8)
-            l |= PixelOdd;
-         if (i & 4)
-            l |= PixelOdd << 8;
-         if (i & 2)
-            l |= PixelOdd << 16;
-         if (i & 1)
-            l |= PixelOdd << 24;
-#else
+#if defined(MSB_FIRST)
          if (i & 8)
             h |= (PixelOdd << 24);
          if (i & 4)
@@ -302,30 +285,30 @@ bool S9xInitGFX()
             l |= (PixelOdd << 8);
          if (i & 1)
             l |= PixelOdd;
-#endif
-
-         odd_high[bitshift][i] = h;
-         odd_low[bitshift][i] = l;
-         h = l = 0;
-
-#if defined(LSB_FIRST)
+#else
          if (i & 8)
-            h |= PixelEven;
+            h |= PixelOdd;
          if (i & 4)
-            h |= PixelEven << 8;
+            h |= PixelOdd << 8;
          if (i & 2)
-            h |= PixelEven << 16;
+            h |= PixelOdd << 16;
          if (i & 1)
-            h |= PixelEven << 24;
+            h |= PixelOdd << 24;
          if (i & 8)
-            l |= PixelEven;
+            l |= PixelOdd;
          if (i & 4)
-            l |= PixelEven << 8;
+            l |= PixelOdd << 8;
          if (i & 2)
-            l |= PixelEven << 16;
+            l |= PixelOdd << 16;
          if (i & 1)
-            l |= PixelEven << 24;
-#else
+            l |= PixelOdd << 24;
+#endif
+
+         odd_high[bitshift][i] = h;
+         odd_low[bitshift][i] = l;
+         h = l = 0;
+
+#if defined(MSB_FIRST)
          if (i & 8)
             h |= (PixelEven << 24);
          if (i & 4)
@@ -342,6 +325,23 @@ bool S9xInitGFX()
             l |= (PixelEven << 8);
          if (i & 1)
             l |= PixelEven;
+#else
+         if (i & 8)
+            h |= PixelEven;
+         if (i & 4)
+            h |= PixelEven << 8;
+         if (i & 2)
+            h |= PixelEven << 16;
+         if (i & 1)
+            h |= PixelEven << 24;
+         if (i & 8)
+            l |= PixelEven;
+         if (i & 4)
+            l |= PixelEven << 8;
+         if (i & 2)
+            l |= PixelEven << 16;
+         if (i & 1)
+            l |= PixelEven << 24;
 #endif
 
          even_high[bitshift][i] = h;
diff --git a/source/gfx.h b/source/gfx.h
index 98e22ae..c13620a 100644
--- a/source/gfx.h
+++ b/source/gfx.h
@@ -239,15 +239,9 @@ extern uint8_t mul_brightness [16][32];
 #define READ_2BYTES(s) (*(uint16_t *) (s))
 #define WRITE_2BYTES(s, d) *(uint16_t *) (s) = (d)
 #else
-#ifdef LSB_FIRST
 #define READ_2BYTES(s) (*(uint8_t *) (s) | (*((uint8_t *) (s) + 1) << 8))
 #define WRITE_2BYTES(s, d) *(uint8_t *) (s) = (d), \
             *((uint8_t *) (s) + 1) = (d) >> 8
-#else  // else MSB_FISRT
-#define READ_2BYTES(s) (*(uint8_t *) (s) | (*((uint8_t *) (s) + 1) << 8))
-#define WRITE_2BYTES(s, d) *(uint8_t *) (s) = (d), \
-            *((uint8_t *) (s) + 1) = (d) >> 8
-#endif // LSB_FIRST
 #endif // i386
 
 #define SUB_SCREEN_DEPTH 0
diff --git a/source/globals.c b/source/globals.c
index 0625df8..d7fa677 100644
--- a/source/globals.c
+++ b/source/globals.c
@@ -271,19 +271,19 @@ int NoiseFreq [32] =
 
 uint32_t HeadMask [4] =
 {
-#ifdef LSB_FIRST
-   0xffffffff, 0xffffff00, 0xffff0000, 0xff000000
-#else
+#ifdef MSB_FIRST
    0xffffffff, 0x00ffffff, 0x0000ffff, 0x000000ff
+#else
+   0xffffffff, 0xffffff00, 0xffff0000, 0xff000000
 #endif
 };
 
 uint32_t TailMask [5] =
 {
-#ifdef LSB_FIRST
-   0x00000000, 0x000000ff, 0x0000ffff, 0x00ffffff, 0xffffffff
-#else
+#ifdef MSB_FIRST
    0x00000000, 0xff000000, 0xffff0000, 0xffffff00, 0xffffffff
+#else
+   0x00000000, 0x000000ff, 0x0000ffff, 0x00ffffff, 0xffffffff
 #endif
 };
 
diff --git a/source/memmap.c b/source/memmap.c
index 951c5bb..12d65ff 100644
--- a/source/memmap.c
+++ b/source/memmap.c
@@ -815,12 +815,12 @@ again:
 #ifdef DETECT_NASTY_FX_INTERLEAVE
       //MK: Damn. YI trips a BRK currently. Maybe even on a real cart.
 
-#ifdef LSB_FIRST
+#ifdef MSB_FIRST
       if (strncmp((char*) &ROM [0x7fc0], "YOSHI'S ISLAND", 14) == 0
-            && (*(uint16_t*)&ROM[0x7FDE]) == 57611 && ROM[0x10002] == 0xA9)
+            && (ROM[0x7FDE] + (ROM[0x7FDF] << 8)) == 57611 && ROM[0x10002] == 0xA9)
 #else
       if (strncmp((char*) &ROM [0x7fc0], "YOSHI'S ISLAND", 14) == 0
-            && (ROM[0x7FDE] + (ROM[0x7FDF] << 8)) == 57611 && ROM[0x10002] == 0xA9)
+            && (*(uint16_t*)&ROM[0x7FDE]) == 57611 && ROM[0x10002] == 0xA9)
 #endif
       {
          Interleaved = true;
diff --git a/source/port.h b/source/port.h
index 264425f..9dd2229 100644
--- a/source/port.h
+++ b/source/port.h
@@ -202,16 +202,13 @@ void MixSound(void);
 
 #if defined(__i386__) || defined(__i486__) || defined(__i586__) || \
     defined(__WIN32__) || defined(__alpha__)
-#define LSB_FIRST
 #define FAST_LSB_WORD_ACCESS
 #elif defined(__MIPSEL__)
-#define LSB_FIRST
 // On little-endian MIPS, a 16-bit word can be read directly from an address
 // only if it's aligned.
 #define FAST_ALIGNED_LSB_WORD_ACCESS
 #else
 //#define MSB_FIRST
-#define LSB_FIRST
 //#define FAST_LSB_WORD_ACCESS
 #endif
 
diff --git a/source/spc700.h b/source/spc700.h
index 9ddb72f..5c0bb14 100644
--- a/source/spc700.h
+++ b/source/spc700.h
@@ -135,15 +135,15 @@
 
 typedef union
 {
-#ifdef LSB_FIRST
+#ifdef MSB_FIRST
    struct
    {
-      uint8_t A, Y;
+      uint8_t Y, A;
    } B;
 #else
    struct
    {
-      uint8_t Y, A;
+      uint8_t A, Y;
    } B;
 #endif
    uint16_t W;
diff --git a/source/srtc.c b/source/srtc.c
index 5073d35..2d63e96 100644
--- a/source/srtc.c
+++ b/source/srtc.c
@@ -518,10 +518,7 @@ void S9xSRTCPreSaveState()
       Memory.SRAM [s + 3 + MAX_RTC_INDEX] = rtc.index;
       Memory.SRAM [s + 4 + MAX_RTC_INDEX] = rtc.mode;
 
-#ifdef LSB_FIRST
-      // memmove converted: Different mallocs [Neb]
-      memcpy(&Memory.SRAM [s + 5 + MAX_RTC_INDEX], &rtc.system_timestamp, 8);
-#else
+#ifdef MSB_FIRST
       SRAM [s + 5  + MAX_RTC_INDEX] = (uint8_t)(rtc.system_timestamp >>  0);
       SRAM [s + 6  + MAX_RTC_INDEX] = (uint8_t)(rtc.system_timestamp >>  8);
       SRAM [s + 7  + MAX_RTC_INDEX] = (uint8_t)(rtc.system_timestamp >> 16);
@@ -530,6 +527,9 @@ void S9xSRTCPreSaveState()
       SRAM [s + 10 + MAX_RTC_INDEX] = (uint8_t)(rtc.system_timestamp >> 40);
       SRAM [s + 11 + MAX_RTC_INDEX] = (uint8_t)(rtc.system_timestamp >> 48);
       SRAM [s + 12 + MAX_RTC_INDEX] = (uint8_t)(rtc.system_timestamp >> 56);
+#else
+      // memmove converted: Different mallocs [Neb]
+      memcpy(&Memory.SRAM [s + 5 + MAX_RTC_INDEX], &rtc.system_timestamp, 8);
 #endif
    }
 }
@@ -550,10 +550,7 @@ void S9xSRTCPostLoadState()
       rtc.index = Memory.SRAM [s + 3 + MAX_RTC_INDEX];
       rtc.mode = Memory.SRAM [s + 4 + MAX_RTC_INDEX];
 
-#ifdef LSB_FIRST
-      // memmove converted: Different mallocs [Neb]
-      memcpy(&rtc.system_timestamp, &Memory.SRAM [s + 5 + MAX_RTC_INDEX], 8);
-#else
+#ifdef MSB_FIRST
       rtc.system_timestamp |= (SRAM [s +  5 + MAX_RTC_INDEX] <<  0);
       rtc.system_timestamp |= (SRAM [s +  6 + MAX_RTC_INDEX] <<  8);
       rtc.system_timestamp |= (SRAM [s +  7 + MAX_RTC_INDEX] << 16);
@@ -562,6 +559,9 @@ void S9xSRTCPostLoadState()
       rtc.system_timestamp |= (SRAM [s + 10 + MAX_RTC_INDEX] << 40);
       rtc.system_timestamp |= (SRAM [s + 11 + MAX_RTC_INDEX] << 48);
       rtc.system_timestamp |= (SRAM [s + 12 + MAX_RTC_INDEX] << 56);
+#else
+      // memmove converted: Different mallocs [Neb]
+      memcpy(&rtc.system_timestamp, &Memory.SRAM [s + 5 + MAX_RTC_INDEX], 8);
 #endif
       S9xUpdateSrtcTime();
    }