Extend blitting support for all <8bpp formats

src/video/SDL_blit_0.c

@@ -26,8 +26,11 @@
 
 /* Functions to blit from bitmaps to other surfaces */
 
-static void BlitBto1(SDL_BlitInfo *info)
+SDL_FORCE_INLINE void BlitBto1(SDL_BlitInfo *info, const Uint32 srcbpp)
 {
+    const Uint32 mask = (1 << srcbpp) - 1;
+    const Uint32 align = (8 / srcbpp) - 1;
+
     int c;
     int width, height;
     Uint8 *src, *map, *dst;
@@ -41,22 +44,28 @@ static void BlitBto1(SDL_BlitInfo *info)
     dst = info->dst;
     dstskip = info->dst_skip;
     map = info->table;
+
+    if (srcbpp == 4)
+        srcskip += width - (width + 1) / 2;
+    else if (srcbpp == 2)
+        srcskip += width - (width + 3) / 4;
+    else if (srcbpp == 1)
         srcskip += width - (width + 7) / 8;
 
     if (map) {
-        if (info->src_fmt->format == SDL_PIXELFORMAT_INDEX1LSB) {
+        if (SDL_PIXELORDER(info->src_fmt->format) == SDL_BITMAPORDER_4321) {
             while (height--) {
                 Uint8 byte = 0, bit;
                 for (c = 0; c < width; ++c) {
-                    if (!(c & 7)) {
+                    if (!(c & align)) {
                         byte = *src++;
                     }
-                    bit = (byte & 0x01);
+                    bit = (byte & mask);
                     if (1) {
                         *dst = map[bit];
                     }
                     dst++;
-                    byte >>= 1;
+                    byte >>= srcbpp;
                 }
                 src += srcskip;
                 dst += dstskip;
@@ -65,34 +74,34 @@ static void BlitBto1(SDL_BlitInfo *info)
             while (height--) {
                 Uint8 byte = 0, bit;
                 for (c = 0; c < width; ++c) {
-                    if (!(c & 7)) {
+                    if (!(c & align)) {
                         byte = *src++;
                     }
-                    bit = (byte & 0x80) >> 7;
+                    bit = (byte >> (8 - srcbpp)) & mask;
                     if (1) {
                         *dst = map[bit];
                     }
                     dst++;
-                    byte <<= 1;
+                    byte <<= srcbpp;
                 }
                 src += srcskip;
                 dst += dstskip;
             }
         }
     } else {
-        if (info->src_fmt->format == SDL_PIXELFORMAT_INDEX1LSB) {
+        if (SDL_PIXELORDER(info->src_fmt->format) == SDL_BITMAPORDER_4321) {
             while (height--) {
                 Uint8 byte = 0, bit;
                 for (c = 0; c < width; ++c) {
-                    if (!(c & 7)) {
+                    if (!(c & align)) {
                         byte = *src++;
                     }
-                    bit = (byte & 0x01);
+                    bit = (byte & mask);
                     if (1) {
                         *dst = bit;
                     }
                     dst++;
-                    byte >>= 1;
+                    byte >>= srcbpp;
                 }
                 src += srcskip;
                 dst += dstskip;
@@ -101,15 +110,15 @@ static void BlitBto1(SDL_BlitInfo *info)
             while (height--) {
                 Uint8 byte = 0, bit;
                 for (c = 0; c < width; ++c) {
-                    if (!(c & 7)) {
+                    if (!(c & align)) {
                         byte = *src++;
                     }
-                    bit = (byte & 0x80) >> 7;
+                    bit = (byte >> (8 - srcbpp)) & mask;
                     if (1) {
                         *dst = bit;
                     }
                     dst++;
-                    byte <<= 1;
+                    byte <<= srcbpp;
                 }
                 src += srcskip;
                 dst += dstskip;
@@ -118,8 +127,11 @@ static void BlitBto1(SDL_BlitInfo *info)
     }
 }
 
-static void BlitBto2(SDL_BlitInfo *info)
+SDL_FORCE_INLINE void BlitBto2(SDL_BlitInfo *info, const Uint32 srcbpp)
 {
+    const Uint32 mask = (1 << srcbpp) - 1;
+    const Uint32 align = (8 / srcbpp) - 1;
+
     int c;
     int width, height;
     Uint8 *src;
@@ -134,20 +146,26 @@ static void BlitBto2(SDL_BlitInfo *info)
     dst = (Uint16 *)info->dst;
     dstskip = info->dst_skip / 2;
     map = (Uint16 *)info->table;
+
+    if (srcbpp == 4)
+        srcskip += width - (width + 1) / 2;
+    else if (srcbpp == 2)
+        srcskip += width - (width + 3) / 4;
+    else if (srcbpp == 1)
         srcskip += width - (width + 7) / 8;
 
-    if (info->src_fmt->format == SDL_PIXELFORMAT_INDEX1LSB) {
+    if (SDL_PIXELORDER(info->src_fmt->format) == SDL_BITMAPORDER_4321) {
         while (height--) {
             Uint8 byte = 0, bit;
             for (c = 0; c < width; ++c) {
-                if (!(c & 7)) {
+                if (!(c & align)) {
                     byte = *src++;
                 }
-                bit = (byte & 0x01);
+                bit = (byte & mask);
                 if (1) {
                     *dst = map[bit];
                 }
-                byte >>= 1;
+                byte >>= srcbpp;
                 dst++;
             }
             src += srcskip;
@@ -157,14 +175,14 @@ static void BlitBto2(SDL_BlitInfo *info)
         while (height--) {
             Uint8 byte = 0, bit;
             for (c = 0; c < width; ++c) {
-                if (!(c & 7)) {
+                if (!(c & align)) {
                     byte = *src++;
                 }
-                bit = (byte & 0x80) >> 7;
+                bit = (byte >> (8 - srcbpp)) & mask;
                 if (1) {
                     *dst = map[bit];
                 }
-                byte <<= 1;
+                byte <<= srcbpp;
                 dst++;
             }
             src += srcskip;
@@ -173,8 +191,11 @@ static void BlitBto2(SDL_BlitInfo *info)
     }
 }
 
-static void BlitBto3(SDL_BlitInfo *info)
+SDL_FORCE_INLINE void BlitBto3(SDL_BlitInfo *info, const Uint32 srcbpp)
 {
+    const Uint32 mask = (1 << srcbpp) - 1;
+    const Uint32 align = (8 / srcbpp) - 1;
+
     int c, o;
     int width, height;
     Uint8 *src, *map, *dst;
@@ -188,23 +209,29 @@ static void BlitBto3(SDL_BlitInfo *info)
     dst = info->dst;
     dstskip = info->dst_skip;
     map = info->table;
+
+    if (srcbpp == 4)
+        srcskip += width - (width + 1) / 2;
+    else if (srcbpp == 2)
+        srcskip += width - (width + 3) / 4;
+    else if (srcbpp == 1)
         srcskip += width - (width + 7) / 8;
 
-    if (info->src_fmt->format == SDL_PIXELFORMAT_INDEX1LSB) {
+    if (SDL_PIXELORDER(info->src_fmt->format) == SDL_BITMAPORDER_4321) {
         while (height--) {
             Uint8 byte = 0, bit;
             for (c = 0; c < width; ++c) {
-                if (!(c & 7)) {
+                if (!(c & align)) {
                     byte = *src++;
                 }
-                bit = (byte & 0x01);
+                bit = (byte & mask);
                 if (1) {
                     o = bit * 4;
                     dst[0] = map[o++];
                     dst[1] = map[o++];
                     dst[2] = map[o++];
                 }
-                byte >>= 1;
+                byte >>= srcbpp;
                 dst += 3;
             }
             src += srcskip;
@@ -214,17 +241,17 @@ static void BlitBto3(SDL_BlitInfo *info)
         while (height--) {
             Uint8 byte = 0, bit;
             for (c = 0; c < width; ++c) {
-                if (!(c & 7)) {
+                if (!(c & align)) {
                     byte = *src++;
                 }
-                bit = (byte & 0x80) >> 7;
+                bit = (byte >> (8 - srcbpp)) & mask;
                 if (1) {
                     o = bit * 4;
                     dst[0] = map[o++];
                     dst[1] = map[o++];
                     dst[2] = map[o++];
                 }
-                byte <<= 1;
+                byte <<= srcbpp;
                 dst += 3;
             }
             src += srcskip;
@@ -233,8 +260,11 @@ static void BlitBto3(SDL_BlitInfo *info)
     }
 }
 
-static void BlitBto4(SDL_BlitInfo *info)
+SDL_FORCE_INLINE void BlitBto4(SDL_BlitInfo *info, const Uint32 srcbpp)
 {
+    const Uint32 mask = (1 << srcbpp) - 1;
+    const Uint32 align = (8 / srcbpp) - 1;
+
     int width, height;
     Uint8 *src;
     Uint32 *map, *dst;
@@ -249,20 +279,26 @@ static void BlitBto4(SDL_BlitInfo *info)
     dst = (Uint32 *)info->dst;
     dstskip = info->dst_skip / 4;
     map = (Uint32 *)info->table;
+
+    if (srcbpp == 4)
+        srcskip += width - (width + 1) / 2;
+    else if (srcbpp == 2)
+        srcskip += width - (width + 3) / 4;
+    else if (srcbpp == 1)
         srcskip += width - (width + 7) / 8;
 
-    if (info->src_fmt->format == SDL_PIXELFORMAT_INDEX1LSB) {
+    if (SDL_PIXELORDER(info->src_fmt->format) == SDL_BITMAPORDER_4321) {
         while (height--) {
             Uint8 byte = 0, bit;
             for (c = 0; c < width; ++c) {
-                if (!(c & 7)) {
+                if (!(c & align)) {
                     byte = *src++;
                 }
-                bit = (byte & 0x01);
+                bit = (byte & mask);
                 if (1) {
                     *dst = map[bit];
                 }
-                byte >>= 1;
+                byte >>= srcbpp;
                 dst++;
             }
             src += srcskip;
@@ -272,14 +308,14 @@ static void BlitBto4(SDL_BlitInfo *info)
         while (height--) {
             Uint8 byte = 0, bit;
             for (c = 0; c < width; ++c) {
-                if (!(c & 7)) {
+                if (!(c & align)) {
                     byte = *src++;
                 }
-                bit = (byte & 0x80) >> 7;
+                bit = (byte >> (8 - srcbpp)) & mask;
                 if (1) {
                     *dst = map[bit];
                 }
-                byte <<= 1;
+                byte <<= srcbpp;
                 dst++;
             }
             src += srcskip;
@@ -288,8 +324,11 @@ static void BlitBto4(SDL_BlitInfo *info)
     }
 }
 
-static void BlitBto1Key(SDL_BlitInfo *info)
+SDL_FORCE_INLINE void BlitBto1Key(SDL_BlitInfo *info, const Uint32 srcbpp)
 {
+    const Uint32 mask = (1 << srcbpp) - 1;
+    const Uint32 align = (8 / srcbpp) - 1;
+
     int width = info->dst_w;
     int height = info->dst_h;
     Uint8 *src = info->src;
@@ -301,22 +340,27 @@ static void BlitBto1Key(SDL_BlitInfo *info)
     int c;
 
     /* Set up some basic variables */
+    if (srcbpp == 4)
+        srcskip += width - (width + 1) / 2;
+    else if (srcbpp == 2)
+        srcskip += width - (width + 3) / 4;
+    else if (srcbpp == 1)
         srcskip += width - (width + 7) / 8;
 
     if (palmap) {
-        if (info->src_fmt->format == SDL_PIXELFORMAT_INDEX1LSB) {
+        if (SDL_PIXELORDER(info->src_fmt->format) == SDL_BITMAPORDER_4321) {
             while (height--) {
                 Uint8 byte = 0, bit;
                 for (c = 0; c < width; ++c) {
-                    if (!(c & 7)) {
+                    if (!(c & align)) {
                         byte = *src++;
                     }
-                    bit = (byte & 0x01);
+                    bit = (byte & mask);
                     if (bit != ckey) {
                         *dst = palmap[bit];
                     }
                     dst++;
-                    byte >>= 1;
+                    byte >>= srcbpp;
                 }
                 src += srcskip;
                 dst += dstskip;
@@ -325,34 +369,34 @@ static void BlitBto1Key(SDL_BlitInfo *info)
             while (height--) {
                 Uint8 byte = 0, bit;
                 for (c = 0; c < width; ++c) {
-                    if (!(c & 7)) {
+                    if (!(c & align)) {
                         byte = *src++;
                     }
-                    bit = (byte & 0x80) >> 7;
+                    bit = (byte >> (8 - srcbpp)) & mask;
                     if (bit != ckey) {
                         *dst = palmap[bit];
                     }
                     dst++;
-                    byte <<= 1;
+                    byte <<= srcbpp;
                 }
                 src += srcskip;
                 dst += dstskip;
             }
         }
     } else {
-        if (info->src_fmt->format == SDL_PIXELFORMAT_INDEX1LSB) {
+        if (SDL_PIXELORDER(info->src_fmt->format) == SDL_BITMAPORDER_4321) {
             while (height--) {
                 Uint8 byte = 0, bit;
                 for (c = 0; c < width; ++c) {
-                    if (!(c & 7)) {
+                    if (!(c & align)) {
                         byte = *src++;
                     }
-                    bit = (byte & 0x01);
+                    bit = (byte & mask);
                     if (bit != ckey) {
                         *dst = bit;
                     }
                     dst++;
-                    byte >>= 1;
+                    byte >>= srcbpp;
                 }
                 src += srcskip;
                 dst += dstskip;
@@ -361,15 +405,15 @@ static void BlitBto1Key(SDL_BlitInfo *info)
             while (height--) {
                 Uint8 byte = 0, bit;
                 for (c = 0; c < width; ++c) {
-                    if (!(c & 7)) {
+                    if (!(c & align)) {
                         byte = *src++;
                     }
-                    bit = (byte & 0x80) >> 7;
+                    bit = (byte >> (8 - srcbpp)) & mask;
                     if (bit != ckey) {
                         *dst = bit;
                     }
                     dst++;
-                    byte <<= 1;
+                    byte <<= srcbpp;
                 }
                 src += srcskip;
                 dst += dstskip;
@@ -378,8 +422,11 @@ static void BlitBto1Key(SDL_BlitInfo *info)
     }
 }
 
-static void BlitBto2Key(SDL_BlitInfo *info)
+SDL_FORCE_INLINE void BlitBto2Key(SDL_BlitInfo *info, const Uint32 srcbpp)
 {
+    const Uint32 mask = (1 << srcbpp) - 1;
+    const Uint32 align = (8 / srcbpp) - 1;
+
     int width = info->dst_w;
     int height = info->dst_h;
     Uint8 *src = info->src;
@@ -391,21 +438,26 @@ static void BlitBto2Key(SDL_BlitInfo *info)
     int c;
 
     /* Set up some basic variables */
+    if (srcbpp == 4)
+        srcskip += width - (width + 1) / 2;
+    else if (srcbpp == 2)
+        srcskip += width - (width + 3) / 4;
+    else if (srcbpp == 1)
         srcskip += width - (width + 7) / 8;
     dstskip /= 2;
 
-    if (info->src_fmt->format == SDL_PIXELFORMAT_INDEX1LSB) {
+    if (SDL_PIXELORDER(info->src_fmt->format) == SDL_BITMAPORDER_4321) {
         while (height--) {
             Uint8 byte = 0, bit;
             for (c = 0; c < width; ++c) {
-                if (!(c & 7)) {
+                if (!(c & align)) {
                     byte = *src++;
                 }
-                bit = (byte & 0x01);
+                bit = (byte & mask);
                 if (bit != ckey) {
                     *dstp = ((Uint16 *)palmap)[bit];
                 }
-                byte >>= 1;
+                byte >>= srcbpp;
                 dstp++;
             }
             src += srcskip;
@@ -415,14 +467,14 @@ static void BlitBto2Key(SDL_BlitInfo *info)
         while (height--) {
             Uint8 byte = 0, bit;
             for (c = 0; c < width; ++c) {
-                if (!(c & 7)) {
+                if (!(c & align)) {
                     byte = *src++;
                 }
-                bit = (byte & 0x80) >> 7;
+                bit = (byte >> (8 - srcbpp)) & mask;
                 if (bit != ckey) {
                     *dstp = ((Uint16 *)palmap)[bit];
                 }
-                byte <<= 1;
+                byte <<= srcbpp;
                 dstp++;
             }
             src += srcskip;
@@ -431,8 +483,11 @@ static void BlitBto2Key(SDL_BlitInfo *info)
     }
 }
 
-static void BlitBto3Key(SDL_BlitInfo *info)
+SDL_FORCE_INLINE void BlitBto3Key(SDL_BlitInfo *info, const Uint32 srcbpp)
 {
+    const Uint32 mask = (1 << srcbpp) - 1;
+    const Uint32 align = (8 / srcbpp) - 1;
+
     int width = info->dst_w;
     int height = info->dst_h;
     Uint8 *src = info->src;
@@ -444,20 +499,25 @@ static void BlitBto3Key(SDL_BlitInfo *info)
     int c;
 
     /* Set up some basic variables */
+    if (srcbpp == 4)
+        srcskip += width - (width + 1) / 2;
+    else if (srcbpp == 2)
+        srcskip += width - (width + 3) / 4;
+    else if (srcbpp == 1)
         srcskip += width - (width + 7) / 8;
 
-    if (info->src_fmt->format == SDL_PIXELFORMAT_INDEX1LSB) {
+    if (SDL_PIXELORDER(info->src_fmt->format) == SDL_BITMAPORDER_4321) {
         while (height--) {
             Uint8 byte = 0, bit;
             for (c = 0; c < width; ++c) {
-                if (!(c & 7)) {
+                if (!(c & align)) {
                     byte = *src++;
                 }
-                bit = (byte & 0x01);
+                bit = (byte & mask);
                 if (bit != ckey) {
                     SDL_memcpy(dst, &palmap[bit * 4], 3);
                 }
-                byte >>= 1;
+                byte >>= srcbpp;
                 dst += 3;
             }
             src += srcskip;
@@ -467,14 +527,14 @@ static void BlitBto3Key(SDL_BlitInfo *info)
         while (height--) {
             Uint8 byte = 0, bit;
             for (c = 0; c < width; ++c) {
-                if (!(c & 7)) {
+                if (!(c & align)) {
                     byte = *src++;
                 }
-                bit = (byte & 0x80) >> 7;
+                bit = (byte >> (8 - srcbpp)) & mask;
                 if (bit != ckey) {
                     SDL_memcpy(dst, &palmap[bit * 4], 3);
                 }
-                byte <<= 1;
+                byte <<= srcbpp;
                 dst += 3;
             }
             src += srcskip;
@@ -483,8 +543,11 @@ static void BlitBto3Key(SDL_BlitInfo *info)
     }
 }
 
-static void BlitBto4Key(SDL_BlitInfo *info)
+SDL_FORCE_INLINE void BlitBto4Key(SDL_BlitInfo *info, const Uint32 srcbpp)
 {
+    const Uint32 mask = (1 << srcbpp) - 1;
+    const Uint32 align = (8 / srcbpp) - 1;
+
     int width = info->dst_w;
     int height = info->dst_h;
     Uint8 *src = info->src;
@@ -496,21 +559,26 @@ static void BlitBto4Key(SDL_BlitInfo *info)
     int c;
 
     /* Set up some basic variables */
+    if (srcbpp == 4)
+        srcskip += width - (width + 1) / 2;
+    else if (srcbpp == 2)
+        srcskip += width - (width + 3) / 4;
+    else if (srcbpp == 1)
         srcskip += width - (width + 7) / 8;
     dstskip /= 4;
 
-    if (info->src_fmt->format == SDL_PIXELFORMAT_INDEX1LSB) {
+    if (SDL_PIXELORDER(info->src_fmt->format) == SDL_BITMAPORDER_4321) {
         while (height--) {
             Uint8 byte = 0, bit;
             for (c = 0; c < width; ++c) {
-                if (!(c & 7)) {
+                if (!(c & align)) {
                     byte = *src++;
                 }
-                bit = (byte & 0x01);
+                bit = (byte & mask);
                 if (bit != ckey) {
                     *dstp = ((Uint32 *)palmap)[bit];
                 }
-                byte >>= 1;
+                byte >>= srcbpp;
                 dstp++;
             }
             src += srcskip;
@@ -520,14 +588,14 @@ static void BlitBto4Key(SDL_BlitInfo *info)
         while (height--) {
             Uint8 byte = 0, bit;
             for (c = 0; c < width; ++c) {
-                if (!(c & 7)) {
+                if (!(c & align)) {
                     byte = *src++;
                 }
-                bit = (byte & 0x80) >> 7;
+                bit = (byte >> (8 - srcbpp)) & mask;
                 if (bit != ckey) {
                     *dstp = ((Uint32 *)palmap)[bit];
                 }
-                byte <<= 1;
+                byte <<= srcbpp;
                 dstp++;
             }
             src += srcskip;
@@ -536,8 +604,11 @@ static void BlitBto4Key(SDL_BlitInfo *info)
     }
 }
 
-static void BlitBtoNAlpha(SDL_BlitInfo *info)
+SDL_FORCE_INLINE void BlitBtoNAlpha(SDL_BlitInfo *info, const Uint32 srcbpp)
 {
+    const Uint32 mask = (1 << srcbpp) - 1;
+    const Uint32 align = (8 / srcbpp) - 1;
+
     int width = info->dst_w;
     int height = info->dst_h;
     Uint8 *src = info->src;
@@ -555,16 +626,21 @@ static void BlitBtoNAlpha(SDL_BlitInfo *info)
 
     /* Set up some basic variables */
     dstbpp = dstfmt->BytesPerPixel;
+    if (srcbpp == 4)
+        srcskip += width - (width + 1) / 2;
+    else if (srcbpp == 2)
+        srcskip += width - (width + 3) / 4;
+    else if (srcbpp == 1)
         srcskip += width - (width + 7) / 8;
 
-    if (info->src_fmt->format == SDL_PIXELFORMAT_INDEX1LSB) {
+    if (SDL_PIXELORDER(info->src_fmt->format) == SDL_BITMAPORDER_4321) {
         while (height--) {
             Uint8 byte = 0, bit;
             for (c = 0; c < width; ++c) {
-                if (!(c & 7)) {
+                if (!(c & align)) {
                     byte = *src++;
                 }
-                bit = (byte & 0x01);
+                bit = (byte & mask);
                 if (1) {
                     sR = srcpal[bit].r;
                     sG = srcpal[bit].g;
@@ -573,7 +649,7 @@ static void BlitBtoNAlpha(SDL_BlitInfo *info)
                     ALPHA_BLEND_RGBA(sR, sG, sB, A, dR, dG, dB, dA);
                     ASSEMBLE_RGBA(dst, dstbpp, dstfmt, dR, dG, dB, dA);
                 }
-                byte >>= 1;
+                byte >>= srcbpp;
                 dst += dstbpp;
             }
             src += srcskip;
@@ -583,10 +659,10 @@ static void BlitBtoNAlpha(SDL_BlitInfo *info)
         while (height--) {
             Uint8 byte = 0, bit;
             for (c = 0; c < width; ++c) {
-                if (!(c & 7)) {
+                if (!(c & align)) {
                     byte = *src++;
                 }
-                bit = (byte & 0x80) >> 7;
+                bit = (byte >> (8 - srcbpp)) & mask;
                 if (1) {
                     sR = srcpal[bit].r;
                     sG = srcpal[bit].g;
@@ -595,7 +671,7 @@ static void BlitBtoNAlpha(SDL_BlitInfo *info)
                     ALPHA_BLEND_RGBA(sR, sG, sB, A, dR, dG, dB, dA);
                     ASSEMBLE_RGBA(dst, dstbpp, dstfmt, dR, dG, dB, dA);
                 }
-                byte <<= 1;
+                byte <<= srcbpp;
                 dst += dstbpp;
             }
             src += srcskip;
@@ -604,8 +680,11 @@ static void BlitBtoNAlpha(SDL_BlitInfo *info)
     }
 }
 
-static void BlitBtoNAlphaKey(SDL_BlitInfo *info)
+SDL_FORCE_INLINE void BlitBtoNAlphaKey(SDL_BlitInfo *info, const Uint32 srcbpp)
 {
+    const Uint32 mask = (1 << srcbpp) - 1;
+    const Uint32 align = (8 / srcbpp) - 1;
+
     int width = info->dst_w;
     int height = info->dst_h;
     Uint8 *src = info->src;
@@ -625,16 +704,21 @@ static void BlitBtoNAlphaKey(SDL_BlitInfo *info)
 
     /* Set up some basic variables */
     dstbpp = dstfmt->BytesPerPixel;
+    if (srcbpp == 4)
+        srcskip += width - (width + 1) / 2;
+    else if (srcbpp == 2)
+        srcskip += width - (width + 3) / 4;
+    else if (srcbpp == 1)
         srcskip += width - (width + 7) / 8;
 
-    if (info->src_fmt->format == SDL_PIXELFORMAT_INDEX1LSB) {
+    if (SDL_PIXELORDER(info->src_fmt->format) == SDL_BITMAPORDER_4321) {
         while (height--) {
             Uint8 byte = 0, bit;
             for (c = 0; c < width; ++c) {
-                if (!(c & 7)) {
+                if (!(c & align)) {
                     byte = *src++;
                 }
-                bit = (byte & 0x01);
+                bit = (byte & mask);
                 if (bit != ckey) {
                     sR = srcpal[bit].r;
                     sG = srcpal[bit].g;
@@ -643,7 +727,7 @@ static void BlitBtoNAlphaKey(SDL_BlitInfo *info)
                     ALPHA_BLEND_RGBA(sR, sG, sB, A, dR, dG, dB, dA);
                     ASSEMBLE_RGBA(dst, dstbpp, dstfmt, dR, dG, dB, dA);
                 }
-                byte >>= 1;
+                byte >>= srcbpp;
                 dst += dstbpp;
             }
             src += srcskip;
@@ -653,10 +737,10 @@ static void BlitBtoNAlphaKey(SDL_BlitInfo *info)
         while (height--) {
             Uint8 byte = 0, bit;
             for (c = 0; c < width; ++c) {
-                if (!(c & 7)) {
+                if (!(c & align)) {
                     byte = *src++;
                 }
-                bit = (byte & 0x80) >> 7;
+                bit = (byte >> (8 - srcbpp)) & mask;
                 if (bit != ckey) {
                     sR = srcpal[bit].r;
                     sG = srcpal[bit].g;
@@ -665,7 +749,7 @@ static void BlitBtoNAlphaKey(SDL_BlitInfo *info)
                     ALPHA_BLEND_RGBA(sR, sG, sB, A, dR, dG, dB, dA);
                     ASSEMBLE_RGBA(dst, dstbpp, dstfmt, dR, dG, dB, dA);
                 }
-                byte <<= 1;
+                byte <<= srcbpp;
                 dst += dstbpp;
             }
             src += srcskip;
@@ -674,115 +758,221 @@ static void BlitBtoNAlphaKey(SDL_BlitInfo *info)
     }
 }
 
-static const SDL_BlitFunc bitmap_blit[] = {
-    (SDL_BlitFunc)NULL, BlitBto1, BlitBto2, BlitBto3, BlitBto4
-};
 
-static const SDL_BlitFunc colorkey_blit[] = {
-    (SDL_BlitFunc)NULL, BlitBto1Key, BlitBto2Key, BlitBto3Key, BlitBto4Key
-};
 
-static void Blit4bto4(SDL_BlitInfo *info)
+static void Blit1bto1(SDL_BlitInfo *info) {
-{
+    BlitBto1(info, 1);
-    int width = info->dst_w;
-    int height = info->dst_h;
-    Uint8 *src = info->src;
-    Uint32 *dst = (Uint32 *)info->dst;
-    int srcskip = info->src_skip;
-    int dstskip = info->dst_skip;
-    Uint32 *map = (Uint32 *)info->table;
-    int c;
-
-    /* Set up some basic variables */
-    srcskip += width - (width + 1) / 2;
-
-    while (height--) {
-        Uint8 byte = 0, bit;
-        for (c = 0; c < width; ++c) {
-            if (!(c & 0x1)) {
-                byte = *src++;
-            }
-            bit = (byte & 0xF0) >> 4;
-            if (1) {
-                *dst = map[bit];
-            }
-            byte <<= 4;
-            dst++;
-        }
-        src += srcskip;
-        dst = (Uint32 *)((Uint8 *)dst + dstskip);
-    }
 }
 
-static void Blit4bto4Key(SDL_BlitInfo *info)
+static void Blit1bto2(SDL_BlitInfo *info) {
-{
+    BlitBto2(info, 1);
-    int width = info->dst_w;
-    int height = info->dst_h;
-    Uint8 *src = info->src;
-    Uint32 *dst = (Uint32 *)info->dst;
-    int srcskip = info->src_skip;
-    int dstskip = info->dst_skip;
-    Uint32 ckey = info->colorkey;
-    Uint32 *map = (Uint32 *)info->table;
-    int c;
-
-    /* Set up some basic variables */
-    srcskip += width - (width + 1) / 2;
-
-    while (height--) {
-        Uint8 byte = 0, bit;
-        for (c = 0; c < width; ++c) {
-            if (!(c & 0x1)) {
-                byte = *src++;
-            }
-            bit = (byte & 0xF0) >> 4;
-            if (bit != ckey) {
-                *dst = map[bit];
-            }
-            byte <<= 4;
-            dst++;
-        }
-        src += srcskip;
-        dst = (Uint32 *)((Uint8 *)dst + dstskip);
-    }
 }
 
+static void Blit1bto3(SDL_BlitInfo *info) {
+    BlitBto3(info, 1);
+}
+
+static void Blit1bto4(SDL_BlitInfo *info) {
+    BlitBto4(info, 1);
+}
+
+static const SDL_BlitFunc bitmap_blit_1b[] = {
+    (SDL_BlitFunc)NULL, Blit1bto1, Blit1bto2, Blit1bto3, Blit1bto4
+};
+
+static void Blit1bto1Key(SDL_BlitInfo *info) {
+    BlitBto1Key(info, 1);
+}
+
+static void Blit1bto2Key(SDL_BlitInfo *info) {
+    BlitBto2Key(info, 1);
+}
+
+static void Blit1bto3Key(SDL_BlitInfo *info) {
+    BlitBto3Key(info, 1);
+}
+
+static void Blit1bto4Key(SDL_BlitInfo *info) {
+    BlitBto4Key(info, 1);
+}
+
+static const SDL_BlitFunc colorkey_blit_1b[] = {
+    (SDL_BlitFunc)NULL, Blit1bto1Key, Blit1bto2Key, Blit1bto3Key, Blit1bto4Key
+};
+
+static void Blit1btoNAlpha(SDL_BlitInfo *info)
+{
+    BlitBtoNAlpha(info, 1);
+}
+
+static void Blit1btoNAlphaKey(SDL_BlitInfo *info)
+{
+    BlitBtoNAlphaKey(info, 1);
+}
+
+
+
+static void Blit2bto1(SDL_BlitInfo *info) {
+    BlitBto1(info, 2);
+}
+
+static void Blit2bto2(SDL_BlitInfo *info) {
+    BlitBto2(info, 2);
+}
+
+static void Blit2bto3(SDL_BlitInfo *info) {
+    BlitBto3(info, 2);
+}
+
+static void Blit2bto4(SDL_BlitInfo *info) {
+    BlitBto4(info, 2);
+}
+
+static const SDL_BlitFunc bitmap_blit_2b[] = {
+    (SDL_BlitFunc)NULL, Blit2bto1, Blit2bto2, Blit2bto3, Blit2bto4
+};
+
+static void Blit2bto1Key(SDL_BlitInfo *info) {
+    BlitBto1Key(info, 2);
+}
+
+static void Blit2bto2Key(SDL_BlitInfo *info) {
+    BlitBto2Key(info, 2);
+}
+
+static void Blit2bto3Key(SDL_BlitInfo *info) {
+    BlitBto3Key(info, 2);
+}
+
+static void Blit2bto4Key(SDL_BlitInfo *info) {
+    BlitBto4Key(info, 2);
+}
+
+static const SDL_BlitFunc colorkey_blit_2b[] = {
+    (SDL_BlitFunc)NULL, Blit2bto1Key, Blit2bto2Key, Blit2bto3Key, Blit2bto4Key
+};
+
+static void Blit2btoNAlpha(SDL_BlitInfo *info)
+{
+    BlitBtoNAlpha(info, 2);
+}
+
+static void Blit2btoNAlphaKey(SDL_BlitInfo *info)
+{
+    BlitBtoNAlphaKey(info, 2);
+}
+
+
+
+static void Blit4bto1(SDL_BlitInfo *info) {
+    BlitBto1(info, 4);
+}
+
+static void Blit4bto2(SDL_BlitInfo *info) {
+    BlitBto2(info, 4);
+}
+
+static void Blit4bto3(SDL_BlitInfo *info) {
+    BlitBto3(info, 4);
+}
+
+static void Blit4bto4(SDL_BlitInfo *info) {
+    BlitBto4(info, 4);
+}
+
+static const SDL_BlitFunc bitmap_blit_4b[] = {
+    (SDL_BlitFunc)NULL, Blit4bto1, Blit4bto2, Blit4bto3, Blit4bto4
+};
+
+static void Blit4bto1Key(SDL_BlitInfo *info) {
+    BlitBto1Key(info, 4);
+}
+
+static void Blit4bto2Key(SDL_BlitInfo *info) {
+    BlitBto2Key(info, 4);
+}
+
+static void Blit4bto3Key(SDL_BlitInfo *info) {
+    BlitBto3Key(info, 4);
+}
+
+static void Blit4bto4Key(SDL_BlitInfo *info) {
+    BlitBto4Key(info, 4);
+}
+
+static const SDL_BlitFunc colorkey_blit_4b[] = {
+    (SDL_BlitFunc)NULL, Blit4bto1Key, Blit4bto2Key, Blit4bto3Key, Blit4bto4Key
+};
+
+static void Blit4btoNAlpha(SDL_BlitInfo *info)
+{
+    BlitBtoNAlpha(info, 4);
+}
+
+static void Blit4btoNAlphaKey(SDL_BlitInfo *info)
+{
+    BlitBtoNAlphaKey(info, 4);
+}
+
+
+
 SDL_BlitFunc SDL_CalculateBlit0(SDL_Surface *surface)
 {
     int which;
 
-    /* 4bits to 32bits */
-    if (surface->format->format == SDL_PIXELFORMAT_INDEX4MSB) {
-        if (surface->map->dst->format->BytesPerPixel == 4) {
-            switch (surface->map->info.flags & ~SDL_COPY_RLE_MASK) {
-            case 0:
-                return Blit4bto4;
-
-            case SDL_COPY_COLORKEY:
-                return Blit4bto4Key;
-            }
-        }
-        return NULL;
-    }
-
-    if (surface->format->format == SDL_PIXELFORMAT_INDEX1MSB) {
     if (surface->map->dst->format->BitsPerPixel < 8) {
         which = 0;
     } else {
         which = surface->map->dst->format->BytesPerPixel;
     }
+
+    if (SDL_PIXELTYPE(surface->format->format) == SDL_PIXELTYPE_INDEX1) {
         switch (surface->map->info.flags & ~SDL_COPY_RLE_MASK) {
         case 0:
-            return bitmap_blit[which];
+            return bitmap_blit_1b[which];
 
         case SDL_COPY_COLORKEY:
-            return colorkey_blit[which];
+            return colorkey_blit_1b[which];
 
         case SDL_COPY_MODULATE_ALPHA | SDL_COPY_BLEND:
-            return which >= 2 ? BlitBtoNAlpha : (SDL_BlitFunc)NULL;
+            return which >= 2 ? Blit1btoNAlpha : (SDL_BlitFunc)NULL;
 
         case SDL_COPY_COLORKEY | SDL_COPY_MODULATE_ALPHA | SDL_COPY_BLEND:
-            return which >= 2 ? BlitBtoNAlphaKey : (SDL_BlitFunc)NULL;
+            return which >= 2 ? Blit1btoNAlphaKey : (SDL_BlitFunc)NULL;
+        }
+        return NULL;
+    }
+
+    if (SDL_PIXELTYPE(surface->format->format) == SDL_PIXELTYPE_INDEX2) {
+        switch (surface->map->info.flags & ~SDL_COPY_RLE_MASK) {
+        case 0:
+            return bitmap_blit_2b[which];
+
+        case SDL_COPY_COLORKEY:
+            return colorkey_blit_2b[which];
+
+        case SDL_COPY_MODULATE_ALPHA | SDL_COPY_BLEND:
+            return which >= 2 ? Blit2btoNAlpha : (SDL_BlitFunc)NULL;
+
+        case SDL_COPY_COLORKEY | SDL_COPY_MODULATE_ALPHA | SDL_COPY_BLEND:
+            return which >= 2 ? Blit2btoNAlphaKey : (SDL_BlitFunc)NULL;
+        }
+        return NULL;
+    }
+
+    if (SDL_PIXELTYPE(surface->format->format) == SDL_PIXELTYPE_INDEX4) {
+        switch (surface->map->info.flags & ~SDL_COPY_RLE_MASK) {
+        case 0:
+            return bitmap_blit_4b[which];
+
+        case SDL_COPY_COLORKEY:
+            return colorkey_blit_4b[which];
+
+        case SDL_COPY_MODULATE_ALPHA | SDL_COPY_BLEND:
+            return which >= 2 ? Blit4btoNAlpha : (SDL_BlitFunc)NULL;
+
+        case SDL_COPY_COLORKEY | SDL_COPY_MODULATE_ALPHA | SDL_COPY_BLEND:
+            return which >= 2 ? Blit4btoNAlphaKey : (SDL_BlitFunc)NULL;
         }
         return NULL;
     }