Support saving 1- and 4-bit MSB indexed surfaces as BMP

include/SDL3/SDL_surface.h

@@ -594,10 +594,10 @@ extern SDL_DECLSPEC SDL_Surface * SDLCALL SDL_LoadBMP(const char *file);
 /**
  * Save a surface to a seekable SDL data stream in BMP format.
  *
- * Surfaces with a 24-bit, 32-bit and paletted 8-bit format get saved in the
+ * Surfaces with a 24-bit, 32-bit and paletted format get saved in the
  * BMP directly. Other RGB formats with 8-bit or higher get converted to a
  * 24-bit surface or, if they have an alpha mask or a colorkey, to a 32-bit
- * surface before they are saved. YUV and paletted 1-bit and 4-bit formats are
+ * surface before they are saved. YUV and LSB paletted 1-bit and 4-bit formats are
  * not supported.
  *
  * \param surface the SDL_Surface structure containing the image to be saved.
@@ -620,10 +620,10 @@ extern SDL_DECLSPEC bool SDLCALL SDL_SaveBMP_IO(SDL_Surface *surface, SDL_IOStre
 /**
  * Save a surface to a file in BMP format.
  *
- * Surfaces with a 24-bit, 32-bit and paletted 8-bit format get saved in the
+ * Surfaces with a 24-bit, 32-bit, paletted formats get saved in the
  * BMP directly. Other RGB formats with 8-bit or higher get converted to a
  * 24-bit surface or, if they have an alpha mask or a colorkey, to a 32-bit
- * surface before they are saved. YUV and paletted 1-bit and 4-bit formats are
+ * surface before they are saved. YUV and LSB paletted 1-bit and 4-bit formats are
  * not supported.
  *
  * \param surface the SDL_Surface structure containing the image to be saved.

src/video/SDL_bmp.c

@@ -638,11 +638,17 @@ static bool InitBMPSaveState(BMPSaveState *state, SDL_Surface *surface)
     if (surface->palette && !state->save32bit) {
         if (SDL_BITSPERPIXEL(surface->format) == 8) {
             state->intermediate_surface = surface;
+        } else if (surface->format == SDL_PIXELFORMAT_INDEX1MSB || surface->format == SDL_PIXELFORMAT_INDEX4MSB) {
+            state->intermediate_surface = surface;
         } else {
             SDL_SetError("%u bpp BMP files not supported",
                          SDL_BITSPERPIXEL(surface->format));
             goto error;
         }
+    } else if (surface->format == SDL_PIXELFORMAT_INDEX1MSB || surface->format == SDL_PIXELFORMAT_INDEX4MSB) {
+        SDL_SetError("%u bpp BMP files require a palette",
+                     SDL_BITSPERPIXEL(surface->format));
+        goto error;
     } else if ((surface->format == SDL_PIXELFORMAT_BGR24 && !state->save32bit) ||
                (surface->format == SDL_PIXELFORMAT_BGRA32 && state->save32bit)) {
         state->intermediate_surface = surface;
@@ -726,7 +732,7 @@ static bool SDL_SaveBMP_IO_Internal(BMPSaveState *state, SDL_IOStream *dst, bool
     Uint32 bV5Reserved = 0;
 
     if (SDL_LockSurface(state->intermediate_surface)) {
-        const size_t bw = state->intermediate_surface->w * state->intermediate_surface->fmt->bytes_per_pixel;
+        const size_t bw = (state->intermediate_surface->w * SDL_BITSPERPIXEL(state->intermediate_surface->format) + 7) / 8;
 
         // Set the BMP file header values
         bfSize = 0; // We'll write this when we're done

src/video/SDL_surface.c

@@ -2781,9 +2781,11 @@ Uint32 SDL_MapSurfaceRGBA(SDL_Surface *surface, Uint8 r, Uint8 g, Uint8 b, Uint8
 bool SDL_ReadSurfacePixel(SDL_Surface *surface, int x, int y, Uint8 *r, Uint8 *g, Uint8 *b, Uint8 *a)
 {
     Uint32 pixel = 0;
+    Uint32 mask = 0xffffffff;
+    Uint32 shift = 0;
     size_t bytes_per_pixel;
     Uint8 unused;
-    Uint8 *p;
+    const Uint8 *p;
     bool result = false;
 
     if (r) {
@@ -2830,7 +2832,53 @@ bool SDL_ReadSurfacePixel(SDL_Surface *surface, int x, int y, Uint8 *r, Uint8 *g
         }
     }
 
-    p = (Uint8 *)surface->pixels + y * surface->pitch + x * bytes_per_pixel;
+    if (bytes_per_pixel == 0) {
+        switch (surface->format) {
+        case SDL_PIXELFORMAT_INDEX4MSB:
+            p = (Uint8 *)surface->pixels + y * surface->pitch + x / 2;
+            shift = 4 * (1 - (x & 0x1));
+            mask = 0xf << shift;
+            bytes_per_pixel = 1;
+            break;
+        case SDL_PIXELFORMAT_INDEX4LSB:
+            p = (Uint8 *)surface->pixels + y * surface->pitch + x / 2;
+            shift = 4 * (x & 0x1);
+            mask = 0xf << shift;
+            bytes_per_pixel = 1;
+            break;
+        case SDL_PIXELFORMAT_INDEX2MSB:
+            p = (Uint8 *)surface->pixels + y * surface->pitch + x / 4;
+            shift = 2 * (3 - (x & 0x3));
+            mask = 0x3 << shift;
+            bytes_per_pixel = 1;
+            break;
+        case SDL_PIXELFORMAT_INDEX2LSB:
+            p = (Uint8 *)surface->pixels + y * surface->pitch + x / 4;
+            shift = 2 * (x & 0x3);
+            mask = 0x3 << shift;
+            bytes_per_pixel = 1;
+            break;
+        case SDL_PIXELFORMAT_INDEX1MSB:
+            p = (Uint8 *)surface->pixels + y * surface->pitch + x / 8;
+            shift = 7 - (x & 0x7);
+            mask = 0x1 << shift;
+            bytes_per_pixel = 1;
+            break;
+        case SDL_PIXELFORMAT_INDEX1LSB:
+            p = (Uint8 *)surface->pixels + y * surface->pitch + x / 8;
+            shift = x & 0x7;
+            mask = 0x1 << shift;
+            bytes_per_pixel = 1;
+            break;
+        default:
+            if (SDL_MUSTLOCK(surface)) {
+                SDL_UnlockSurface(surface);
+            }
+            return SDL_SetError("Unsupported format with 0 bpp");
+        }
+    } else {
+        p = (Uint8 *)surface->pixels + y * surface->pitch + x * bytes_per_pixel;
+    }
 
     if (bytes_per_pixel <= sizeof(pixel) &&
         !SDL_ISPIXELFORMAT_FOURCC(surface->format) &&
@@ -2842,6 +2890,7 @@ bool SDL_ReadSurfacePixel(SDL_Surface *surface, int x, int y, Uint8 *r, Uint8 *g
 #else
         SDL_memcpy(&pixel, p, bytes_per_pixel);
 #endif
+        pixel = (pixel & mask) >> shift;
         SDL_GetRGBA(pixel, surface->fmt, surface->palette, r, g, b, a);
         result = true;
     } else if (SDL_ISPIXELFORMAT_FOURCC(surface->format)) {