/* 
 * call-seq:
 *    flip(horz, vert)  ->  Surface
 * 
 *  Flips the source surface horizontally (if +horz+ is true), vertically
 *  (if +vert+ is true), or both (if both are true). This operation is
 *  non-destructive; the original image can be perfectly reconstructed by
 *  flipping the resultant image again.
 *
 *  This operation does NOT require SDL_gfx.
 *
 *  A similar effect can (supposedly) be achieved by giving X or Y zoom
 *  factors of -1 to #rotozoom (only if compiled with SDL_gfx 2.0.13 or
 *  greater). Your mileage may vary.
 */
VALUE rbgm_transform_flip(VALUE self, VALUE vhorz, VALUE vvert)
{
  SDL_Surface *surf, *newsurf;
  int xaxis, yaxis;

  int loopx, loopy;
  int pixsize, srcpitch, dstpitch;
  Uint8 *srcpix, *dstpix;

  xaxis = RTEST(vhorz);
  yaxis = RTEST(vvert);

  Data_Get_Struct(self,SDL_Surface,surf);

  /* Borrowed from Pygame: */
  newsurf = newsurf_fromsurf(surf, surf->w, surf->h);
  if(!newsurf)
    return Qnil;

  pixsize = surf->format->BytesPerPixel;
  srcpitch = surf->pitch;
  dstpitch = newsurf->pitch;

  SDL_LockSurface(newsurf);

  srcpix = (Uint8*)surf->pixels;
  dstpix = (Uint8*)newsurf->pixels;

  if(!xaxis)
  {
    if(!yaxis)
    {
      for(loopy = 0; loopy < surf->h; ++loopy)
        memcpy(dstpix+loopy*dstpitch, srcpix+loopy*srcpitch, surf->w*surf->format->BytesPerPixel);
    }
    else
    {
      for(loopy = 0; loopy < surf->h; ++loopy)
        memcpy(dstpix+loopy*dstpitch, srcpix+(surf->h-1-loopy)*srcpitch, surf->w*surf->format->BytesPerPixel);
    }
  }
  else /*if (xaxis)*/
  {
    if(yaxis)
    {
      switch(surf->format->BytesPerPixel)
      {
      case 1:
        for(loopy = 0; loopy < surf->h; ++loopy) {
          Uint8* dst = (Uint8*)(dstpix+loopy*dstpitch);
          Uint8* src = ((Uint8*)(srcpix+(surf->h-1-loopy)*srcpitch)) + surf->w - 1;
          for(loopx = 0; loopx < surf->w; ++loopx)
            *dst++ = *src--;
        }break;
      case 2:
        for(loopy = 0; loopy < surf->h; ++loopy) {
          Uint16* dst = (Uint16*)(dstpix+loopy*dstpitch);
          Uint16* src = ((Uint16*)(srcpix+(surf->h-1-loopy)*srcpitch)) + surf->w - 1;
          for(loopx = 0; loopx < surf->w; ++loopx)
            *dst++ = *src--;
        }break;
      case 4:
        for(loopy = 0; loopy < surf->h; ++loopy) {
          Uint32* dst = (Uint32*)(dstpix+loopy*dstpitch);
          Uint32* src = ((Uint32*)(srcpix+(surf->h-1-loopy)*srcpitch)) + surf->w - 1;
          for(loopx = 0; loopx < surf->w; ++loopx)
            *dst++ = *src--;
        }break;
      case 3:
        for(loopy = 0; loopy < surf->h; ++loopy) {
          Uint8* dst = (Uint8*)(dstpix+loopy*dstpitch);
          Uint8* src = ((Uint8*)(srcpix+(surf->h-1-loopy)*srcpitch)) + surf->w*3 - 3;
          for(loopx = 0; loopx < surf->w; ++loopx)
          {
            dst[0] = src[0]; dst[1] = src[1]; dst[2] = src[2];
            dst += 3;
            src -= 3;
          }
        }break;
      }
    }
    else
    {
      switch(surf->format->BytesPerPixel)
      {
      case 1:
        for(loopy = 0; loopy < surf->h; ++loopy) {
          Uint8* dst = (Uint8*)(dstpix+loopy*dstpitch);
          Uint8* src = ((Uint8*)(srcpix+loopy*srcpitch)) + surf->w - 1;
          for(loopx = 0; loopx < surf->w; ++loopx)
            *dst++ = *src--;
        }break;
      case 2:
        for(loopy = 0; loopy < surf->h; ++loopy) {
          Uint16* dst = (Uint16*)(dstpix+loopy*dstpitch);
          Uint16* src = ((Uint16*)(srcpix+loopy*srcpitch)) + surf->w - 1;
          for(loopx = 0; loopx < surf->w; ++loopx)
            *dst++ = *src--;
        }break;
      case 4:
        for(loopy = 0; loopy < surf->h; ++loopy) {
          Uint32* dst = (Uint32*)(dstpix+loopy*dstpitch);
          Uint32* src = ((Uint32*)(srcpix+loopy*srcpitch)) + surf->w - 1;
          for(loopx = 0; loopx < surf->w; ++loopx)
            *dst++ = *src--;
        }break;
      case 3:
        for(loopy = 0; loopy < surf->h; ++loopy) {
          Uint8* dst = (Uint8*)(dstpix+loopy*dstpitch);
          Uint8* src = ((Uint8*)(srcpix+loopy*srcpitch)) + surf->w*3 - 3;
          for(loopx = 0; loopx < surf->w; ++loopx)
          {
            dst[0] = src[0]; dst[1] = src[1]; dst[2] = src[2];
            dst += 3;
            src -= 3;
          }
        }break;
      }
    }
  }

  SDL_UnlockSurface(newsurf);
  /* Thanks, Pygame :) */


  if(newsurf == NULL)
    rb_raise(eSDLError,"Could not flip surface: %s",SDL_GetError());
  return Data_Wrap_Struct(cSurface,0,SDL_FreeSurface,newsurf);
}