Attempt to optimise the ADD background drawing mode so it's playable on automatic frameskip in games like Super Metroid. It doesn't work well.

See the video for this bug at <http://www.youtube.com/watch?v=sUWjVxAD9Q8>.

3 files changed, 272 insertions, 141 deletions

diff --git a/source/gfx.cpp b/source/gfx.cpp
index f6a6217..decd195 100644
--- a/source/gfx.cpp
+++ b/source/gfx.cpp
@@ -892,8 +892,8 @@ void S9xSetInfoString (const char *string)
 
 inline void SelectTileRenderer (bool8 normal)
 {
-    if (normal)
-    {
+	if (normal)
+	{
 		if (IPPU.HalfWidthPixels)
 		{
 			DrawTilePtr = DrawTile16HalfWidth;
@@ -906,58 +906,51 @@ inline void SelectTileRenderer (bool8 normal)
 			DrawClippedTilePtr = DrawClippedTile16;
 			DrawLargePixelPtr = DrawLargePixel16;
 		}
-    }
-    else
-    {
-		if (GFX.r2131 & 0x80)
-		{
-		    if (GFX.r2131 & 0x40)
-		    {
-				if (GFX.r2130 & 2)
-				{
-				    DrawTilePtr = DrawTile16Sub1_2;
-				    DrawClippedTilePtr = DrawClippedTile16Sub1_2;
-				}
-				else
-				{
-				    // Fixed colour substraction
-				    DrawTilePtr = DrawTile16FixedSub1_2;
-				    DrawClippedTilePtr = DrawClippedTile16FixedSub1_2;
-				}
-				DrawLargePixelPtr = DrawLargePixel16Sub1_2;
-		    }
-		    else
-		    {
-				DrawTilePtr = DrawTile16Sub;
-				DrawClippedTilePtr = DrawClippedTile16Sub;
-				DrawLargePixelPtr = DrawLargePixel16Sub;
-		    }
-		}
-		else
+	}
+	else
+	{
+		switch (GFX.r2131 & 0xC0)
 		{
-		    if (GFX.r2131 & 0x40)
-		    {
-				if (GFX.r2130 & 2)
-				{
-				    DrawTilePtr = DrawTile16Add1_2;
-				    DrawClippedTilePtr = DrawClippedTile16Add1_2;
-				}
-				else
-				{
-				    // Fixed colour addition
-				    DrawTilePtr = DrawTile16FixedAdd1_2;
-				    DrawClippedTilePtr = DrawClippedTile16FixedAdd1_2;
-				}
-				DrawLargePixelPtr = DrawLargePixel16Add1_2;
-		    }
-		    else
-		    {
-				DrawTilePtr = DrawTile16Add;
-				DrawClippedTilePtr = DrawClippedTile16Add;
-				DrawLargePixelPtr = DrawLargePixel16Add;
-		    }
+		case 0x00:
+			DrawTilePtr = DrawTile16Add;
+			DrawClippedTilePtr = DrawClippedTile16Add;
+			DrawLargePixelPtr = DrawLargePixel16Add;
+			break;
+		case 0x40:
+			if (GFX.r2130 & 2)
+			{
+				DrawTilePtr = DrawTile16Add1_2;
+				DrawClippedTilePtr = DrawClippedTile16Add1_2;
+			}
+			else
+			{
+				// Fixed colour addition
+				DrawTilePtr = DrawTile16FixedAdd1_2;
+				DrawClippedTilePtr = DrawClippedTile16FixedAdd1_2;
+			}
+			DrawLargePixelPtr = DrawLargePixel16Add1_2;
+			break;
+		case 0x80:
+			DrawTilePtr = DrawTile16Sub;
+			DrawClippedTilePtr = DrawClippedTile16Sub;
+			DrawLargePixelPtr = DrawLargePixel16Sub;
+			break;
+		case 0xC0:
+			if (GFX.r2130 & 2)
+			{
+				DrawTilePtr = DrawTile16Sub1_2;
+				DrawClippedTilePtr = DrawClippedTile16Sub1_2;
+			}
+			else
+			{
+				// Fixed colour substraction
+				DrawTilePtr = DrawTile16FixedSub1_2;
+				DrawClippedTilePtr = DrawClippedTile16FixedSub1_2;
+			}
+			DrawLargePixelPtr = DrawLargePixel16Sub1_2;
+			break;
 		}
-    }
+	}
 }
 
 void S9xSetupOBJ ()
diff --git a/source/gfx.h b/source/gfx.h
index d6da783..88b2640 100644
--- a/source/gfx.h
+++ b/source/gfx.h
@@ -93,6 +93,31 @@
 #include "port.h"
 #include "snes9x.h"
 
+START_EXTERN_C
+void S9xStartScreenRefresh ();
+void S9xDrawScanLine (uint8 Line);
+void S9xEndScreenRefresh ();
+void S9xSetupOBJ ();
+void S9xUpdateScreen ();
+void RenderLine (uint8 line);
+void S9xBuildDirectColourMaps ();
+
+// External port interface which must be implemented or initialised for each
+// port.
+extern struct SGFX GFX;
+
+bool8 S9xGraphicsInit ();
+void S9xGraphicsDeinit();
+bool8 S9xInitUpdate (void);
+bool8 S9xDeinitUpdate (int Width, int Height, bool8 sixteen_bit);
+void S9xSyncSpeed ();
+
+#ifdef GFX_MULTI_FORMAT
+bool8 S9xSetRenderPixelFormat (int format);
+#endif
+
+END_EXTERN_C
+
 struct SGFX{
     // Initialize these variables
     uint8  *Screen;
@@ -231,11 +256,17 @@ GFX.X2 [((((C1) & RGB_REMOVE_LOW_BITS_MASK) + \
 	  ((C2) & RGB_REMOVE_LOW_BITS_MASK)) >> 1) + \
 	((C1) & (C2) & RGB_LOW_BITS_MASK)]
 #else
-#define COLOR_ADD(C1, C2) \
-(GFX.X2 [((((C1) & RGB_REMOVE_LOW_BITS_MASK) + \
-	  ((C2) & RGB_REMOVE_LOW_BITS_MASK)) >> 1) + \
-	 ((C1) & (C2) & RGB_LOW_BITS_MASK)] | \
- (((C1) ^ (C2)) & RGB_LOW_BITS_MASK))	   
+inline uint16 COLOR_ADD (uint16, uint16);
+
+inline uint16 COLOR_ADD (uint16 C1, uint16 C2)
+{
+	if (C1 == 0)
+		return C2;
+	else if (C2 == 0)
+		return C1;
+	else
+		return GFX.X2 [(((C1 & RGB_REMOVE_LOW_BITS_MASK) + (C2 & RGB_REMOVE_LOW_BITS_MASK)) >> 1) + (C1 & C2 & RGB_LOW_BITS_MASK)] | ((C1 ^ C2) & RGB_LOW_BITS_MASK);
+}
 #endif
 
 #define COLOR_ADD1_2(C1, C2) \
@@ -288,30 +319,5 @@ typedef void (*LargePixelRenderer) (uint32 Tile, uint32 Offset,
 				    uint32 StartPixel, uint32 Pixels,
 				    uint32 StartLine, uint32 LineCount);
 
-START_EXTERN_C
-void S9xStartScreenRefresh ();
-void S9xDrawScanLine (uint8 Line);
-void S9xEndScreenRefresh ();
-void S9xSetupOBJ ();
-void S9xUpdateScreen ();
-void RenderLine (uint8 line);
-void S9xBuildDirectColourMaps ();
-
-// External port interface which must be implemented or initialised for each
-// port.
-extern struct SGFX GFX;
-
-bool8 S9xGraphicsInit ();
-void S9xGraphicsDeinit();
-bool8 S9xInitUpdate (void);
-bool8 S9xDeinitUpdate (int Width, int Height, bool8 sixteen_bit);
-void S9xSyncSpeed ();
-
-#ifdef GFX_MULTI_FORMAT
-bool8 S9xSetRenderPixelFormat (int format);
-#endif
-
-END_EXTERN_C
-
 #endif
 
diff --git a/source/tile.cpp b/source/tile.cpp
index e18182d..653a6d7 100644
--- a/source/tile.cpp
+++ b/source/tile.cpp
@@ -764,15 +764,18 @@ static void WRITE_4PIXELS16_ADD (uint32 Offset, uint8 *Pixels, uint16 *ScreenCol
 	{
 		if (GFX.Z1 > Depth [N] && (Pixel = Pixels[N]))
 		{
-			if (SubDepth [N])
+			switch (SubDepth [N])
 			{
-				if (SubDepth [N] != 1)
-					Screen [N] = COLOR_ADD (ScreenColors [Pixel], Screen [GFX.Delta + N]);
-				else
-					Screen [N] = COLOR_ADD (ScreenColors [Pixel], GFX.FixedColour);
-			}
-			else
+			case 0:
 				Screen [N] = ScreenColors [Pixel];
+				break;
+			case 1:
+				Screen [N] = COLOR_ADD (ScreenColors [Pixel], GFX.FixedColour);
+				break;
+			default:
+				Screen [N] = COLOR_ADD (ScreenColors [Pixel], Screen [GFX.Delta + N]);
+				break;
+			}
 			Depth [N] = GFX.Z2;
 		}
 	}
@@ -789,15 +792,18 @@ static void WRITE_4PIXELS16_FLIPPED_ADD (uint32 Offset, uint8 *Pixels, uint16 *S
 	{
 		if (GFX.Z1 > Depth [N] && (Pixel = Pixels[3 - N]))
 		{
-			if (SubDepth [N])
+			switch (SubDepth [N])
 			{
-				if (SubDepth [N] != 1)
-					Screen [N] = COLOR_ADD (ScreenColors [Pixel], Screen [GFX.Delta + N]);
-				else
-					Screen [N] = COLOR_ADD (ScreenColors [Pixel], GFX.FixedColour);
-			}
-			else
+			case 0:
 				Screen [N] = ScreenColors [Pixel];
+				break;
+			case 1:
+				Screen [N] = COLOR_ADD (ScreenColors [Pixel], GFX.FixedColour);
+				break;
+			default:
+				Screen [N] = COLOR_ADD (ScreenColors [Pixel], Screen [GFX.Delta + N]);
+				break;
+			}
 			Depth [N] = GFX.Z2;
 		}
 	}
@@ -814,15 +820,18 @@ static void WRITE_4PIXELS16_ADD1_2 (uint32 Offset, uint8 *Pixels, uint16 *Screen
 	{
 		if (GFX.Z1 > Depth [N] && (Pixel = Pixels[N]))
 		{
-			if (SubDepth [N])
+			switch (SubDepth [N])
 			{
-				if (SubDepth [N] != 1)
-					Screen [N] = (uint16) (COLOR_ADD1_2 (ScreenColors [Pixel], Screen [GFX.Delta + N]));
-				else
-					Screen [N] = COLOR_ADD (ScreenColors [Pixel], GFX.FixedColour);
-			}
-			else
+			case 0:
 				Screen [N] = ScreenColors [Pixel];
+				break;
+			case 1:
+				Screen [N] = COLOR_ADD (ScreenColors [Pixel], GFX.FixedColour);
+				break;
+			default:
+				Screen [N] = (uint16) (COLOR_ADD1_2 (ScreenColors [Pixel], Screen [GFX.Delta + N]));
+				break;
+			}
 			Depth [N] = GFX.Z2;
 		}
 	}
@@ -839,15 +848,18 @@ static void WRITE_4PIXELS16_FLIPPED_ADD1_2 (uint32 Offset, uint8 *Pixels, uint16
 	{
 		if (GFX.Z1 > Depth [N] && (Pixel = Pixels[3 - N]))
 		{
-			if (SubDepth [N])
+			switch (SubDepth [N])
 			{
-				if (SubDepth [N] != 1)
-					Screen [N] = (uint16) (COLOR_ADD1_2 (ScreenColors [Pixel], Screen [GFX.Delta + N]));
-				else
-					Screen [N] = COLOR_ADD (ScreenColors [Pixel], GFX.FixedColour);
-			}
-			else
+			case 0:
 				Screen [N] = ScreenColors [Pixel];
+				break;
+			case 1:
+				Screen [N] = COLOR_ADD (ScreenColors [Pixel], GFX.FixedColour);
+				break;
+			default:
+				Screen [N] = (uint16) (COLOR_ADD1_2 (ScreenColors [Pixel], Screen [GFX.Delta + N]));
+				break;
+			}
 			Depth [N] = GFX.Z2;
 		}
 	}
@@ -864,15 +876,18 @@ static void WRITE_4PIXELS16_SUB (uint32 Offset, uint8 *Pixels, uint16 *ScreenCol
 	{
 		if (GFX.Z1 > Depth [N] && (Pixel = Pixels[N]))
 		{
-			if (SubDepth [N])
+			switch (SubDepth [N])
 			{
-				if (SubDepth [N] != 1)
-					Screen [N] = (uint16) COLOR_SUB (ScreenColors [Pixel], Screen [GFX.Delta + N]);
-				else
-					Screen [N] = (uint16) COLOR_SUB (ScreenColors [Pixel], GFX.FixedColour);
-			}
-			else
+			case 0:
 				Screen [N] = ScreenColors [Pixel];
+				break;
+			case 1:
+				Screen [N] = (uint16) COLOR_SUB (ScreenColors [Pixel], GFX.FixedColour);
+				break;
+			default:
+				Screen [N] = (uint16) COLOR_SUB (ScreenColors [Pixel], Screen [GFX.Delta + N]);
+				break;
+			}
 			Depth [N] = GFX.Z2;
 		}
 	}
@@ -889,15 +904,18 @@ static void WRITE_4PIXELS16_FLIPPED_SUB (uint32 Offset, uint8 *Pixels, uint16 *S
 	{
 		if (GFX.Z1 > Depth [N] && (Pixel = Pixels[3 - N]))
 		{
-			if (SubDepth [N])
+			switch (SubDepth [N])
 			{
-				if (SubDepth [N] != 1)
-					Screen [N] = (uint16) COLOR_SUB (ScreenColors [Pixel], Screen [GFX.Delta + N]);
-				else
-					Screen [N] = (uint16) COLOR_SUB (ScreenColors [Pixel], GFX.FixedColour);
-			}
-			else
+			case 0:
 				Screen [N] = ScreenColors [Pixel];
+				break;
+			case 1:
+				Screen [N] = (uint16) COLOR_SUB (ScreenColors [Pixel], GFX.FixedColour);
+				break;
+			default:
+				Screen [N] = (uint16) COLOR_SUB (ScreenColors [Pixel], Screen [GFX.Delta + N]);
+				break;
+			}
 			Depth [N] = GFX.Z2;
 		}
 	}
@@ -914,15 +932,18 @@ static void WRITE_4PIXELS16_SUB1_2 (uint32 Offset, uint8 *Pixels, uint16 *Screen
 	{
 		if (GFX.Z1 > Depth [N] && (Pixel = Pixels[N]))
 		{
-			if (SubDepth [N])
+			switch (SubDepth [N])
 			{
-				if (SubDepth [N] != 1)
-					Screen [N] = (uint16) COLOR_SUB1_2 (ScreenColors [Pixel], Screen [GFX.Delta + N]);
-				else
-					Screen [N] = (uint16) COLOR_SUB (ScreenColors [Pixel], GFX.FixedColour);
-			}
-			else
+			case 0:
 				Screen [N] = ScreenColors [Pixel];
+				break;
+			case 1:
+				Screen [N] = (uint16) COLOR_SUB (ScreenColors [Pixel], GFX.FixedColour);
+				break;
+			default:
+				Screen [N] = (uint16) COLOR_SUB1_2 (ScreenColors [Pixel], Screen [GFX.Delta + N]);
+				break;
+			}
 			Depth [N] = GFX.Z2;
 		}
 	}
@@ -939,15 +960,18 @@ static void WRITE_4PIXELS16_FLIPPED_SUB1_2 (uint32 Offset, uint8 *Pixels, uint16
 	{
 		if (GFX.Z1 > Depth [N] && (Pixel = Pixels[3 - N]))
 		{
-			if (SubDepth [N])
+			switch (SubDepth [N])
 			{
-				if (SubDepth [N] != 1)
-					Screen [N] = (uint16) COLOR_SUB1_2 (ScreenColors [Pixel], Screen [GFX.Delta + N]);
-				else
-					Screen [N] = (uint16) COLOR_SUB (ScreenColors [Pixel], GFX.FixedColour);
-			}
-			else
+			case 0:
 				Screen [N] = ScreenColors [Pixel];
+				break;
+			case 1:
+				Screen [N] = (uint16) COLOR_SUB (ScreenColors [Pixel], GFX.FixedColour);
+				break;
+			default:
+				Screen [N] = (uint16) COLOR_SUB1_2 (ScreenColors [Pixel], Screen [GFX.Delta + N]);
+				break;
+			}
 			Depth [N] = GFX.Z2;
 		}
 	}
@@ -957,10 +981,118 @@ static void WRITE_4PIXELS16_FLIPPED_SUB1_2 (uint32 Offset, uint8 *Pixels, uint16
 void DrawTile16Add (uint32 Tile, uint32 Offset, uint32 StartLine,
 		    uint32 LineCount)
 {
-    TILE_PREAMBLE
-    register uint8 *bp;
+	TILE_PREAMBLE
+	register uint8 *bp;
+	uint8  Pixel;
+	uint16 *Screen = (uint16 *) GFX.S + Offset;
+	uint8  *Depth = GFX.ZBuffer + Offset;
+	uint8  *SubDepth = GFX.SubZBuffer + Offset;
 
-    RENDER_TILE(WRITE_4PIXELS16_ADD, WRITE_4PIXELS16_FLIPPED_ADD, 4)
+    switch (Tile & (V_FLIP | H_FLIP))
+    {
+    case 0:
+	bp = pCache + StartLine;
+	for (l = LineCount; l != 0; l--, bp += 8, Screen += GFX.PPL, Depth += GFX.PPL, SubDepth += GFX.PPL)
+	{
+		for (uint8 N = 0; N < 8; N++)
+		{
+			if (GFX.Z1 > Depth [N] && (Pixel = bp[N]))
+			{
+				switch (SubDepth [N])
+				{
+				case 0:
+					Screen [N] = ScreenColors [Pixel];
+					break;
+				case 1:
+					Screen [N] = COLOR_ADD (ScreenColors [Pixel], GFX.FixedColour);
+					break;
+				default:
+					Screen [N] = COLOR_ADD (ScreenColors [Pixel], Screen [GFX.Delta + N]);
+					break;
+				}
+				Depth [N] = GFX.Z2;
+			}
+		}
+	}
+        break;
+    case H_FLIP:
+	bp = pCache + StartLine;
+	for (l = LineCount; l != 0; l--, bp += 8, Screen += GFX.PPL, Depth += GFX.PPL, SubDepth += GFX.PPL)
+	{
+		for (uint8 N = 0; N < 8; N++)
+		{
+			if (GFX.Z1 > Depth [N] && (Pixel = bp[7 - N]))
+			{
+				switch (SubDepth [N])
+				{
+				case 0:
+					Screen [N] = ScreenColors [Pixel];
+					break;
+				case 1:
+					Screen [N] = COLOR_ADD (ScreenColors [Pixel], GFX.FixedColour);
+					break;
+				default:
+					Screen [N] = COLOR_ADD (ScreenColors [Pixel], Screen [GFX.Delta + N]);
+					break;
+				}
+				Depth [N] = GFX.Z2;
+			}
+		}
+	}
+        break;
+    case H_FLIP | V_FLIP:
+	bp = pCache + 56 - StartLine;
+	for (l = LineCount; l != 0; l--, bp -= 8, Screen += GFX.PPL, Depth += GFX.PPL, SubDepth += GFX.PPL)
+	{
+		for (uint8 N = 0; N < 8; N++)
+		{
+			if (GFX.Z1 > Depth [N] && (Pixel = bp[7 - N]))
+			{
+				switch (SubDepth [N])
+				{
+				case 0:
+					Screen [N] = ScreenColors [Pixel];
+					break;
+				case 1:
+					Screen [N] = COLOR_ADD (ScreenColors [Pixel], GFX.FixedColour);
+					break;
+				default:
+					Screen [N] = COLOR_ADD (ScreenColors [Pixel], Screen [GFX.Delta + N]);
+					break;
+				}
+				Depth [N] = GFX.Z2;
+			}
+		}
+	}
+        break;
+    case V_FLIP:
+	bp = pCache + 56 - StartLine;
+	for (l = LineCount; l != 0; l--, bp -= 8, Screen += GFX.PPL, Depth += GFX.PPL, SubDepth += GFX.PPL)
+	{
+		for (uint8 N = 0; N < 8; N++)
+		{
+			if (GFX.Z1 > Depth [N] && (Pixel = bp[N]))
+			{
+				switch (SubDepth [N])
+				{
+				case 0:
+					Screen [N] = ScreenColors [Pixel];
+					break;
+				case 1:
+					Screen [N] = COLOR_ADD (ScreenColors [Pixel], GFX.FixedColour);
+					break;
+				default:
+					Screen [N] = COLOR_ADD (ScreenColors [Pixel], Screen [GFX.Delta + N]);
+					break;
+				}
+				Depth [N] = GFX.Z2;
+			}
+		}
+	}
+        break;
+    default:
+        break;
+    }
 }
 
 void DrawClippedTile16Add (uint32 Tile, uint32 Offset,