--- rxvt-unicode/src/rxvtimg.C	2012/06/07 19:24:02	1.48
+++ rxvt-unicode/src/rxvtimg.C	2012/06/07 20:31:26	1.56
@@ -73,22 +73,27 @@
   // that both 24 and 32 bpp MUST be supported by any screen that supports xrender
   int depth = gdk_pixbuf_get_has_alpha (pb) ? 32 : 24;
 
+  int byte_order = ecb_big_endian () ? MSBFirst : LSBFirst;
+
   XImage xi;
 
   xi.width            = width;
   xi.height           = height;
   xi.xoffset          = 0;
   xi.format           = ZPixmap;
-  xi.byte_order       = MSBFirst; // maybe go for host byte order, because servers are usually local?
-  xi.bitmap_unit      = 32;
-  xi.bitmap_bit_order = MSBFirst;
+  xi.byte_order       = ImageByteOrder (dpy);
+  xi.bitmap_unit      = 0;         //XY only, unused
+  xi.bitmap_bit_order = 0;         //XY only, unused
   xi.bitmap_pad       = BitmapPad (dpy);
   xi.depth            = depth;
   xi.bytes_per_line   = 0;
-  xi.bits_per_pixel   = 32;
-  xi.red_mask         = 0x00ff0000;
-  xi.green_mask       = 0x0000ff00;
-  xi.blue_mask        = 0x000000ff;
+  xi.bits_per_pixel   = 32;         //Z only
+  xi.red_mask         = 0x00000000; //Z only, unused
+  xi.green_mask       = 0x00000000; //Z only, unused
+  xi.blue_mask        = 0x00000000; //Z only, unused
+  xi.obdata           = 0;          // probably unused
+
+  bool byte_order_mismatch = byte_order != xi.byte_order;
 
   if (!XInitImage (&xi))
     rxvt_fatal ("unable to initialise ximage, please report.\n");
@@ -106,31 +111,41 @@
 
   for (int y = 0; y < height; y++)
     {
-      unsigned char r, g, b, a;
-      unsigned char *data = row;
+      unsigned char *src = row;
+      uint32_t      *dst = (uint32_t *)line;
 
       if (depth == 24)
         for (int x = 0; x < width; x++)
           {
-            asm volatile("nop");
-            r = *data++;
-            g = *data++;
-            b = *data++;
-            *line++ = 0;
-            *line++ = r;
-            *line++ = g;
-            *line++ = b;
-            asm volatile("nop");
+            uint8_t r = *src++;
+            uint8_t g = *src++;
+            uint8_t b = *src++;
+
+            uint32_t v = (r << 16) | (g << 8) | b;
+            
+            if (ecb_big_endian () ? !byte_order_mismatch : byte_order_mismatch)
+              v = ecb_bswap32 (v);
+
+            *dst++ = v;
           }
       else
         for (int x = 0; x < width; x++)
           {
-            uint32_t v = *(uint32_t *)data; data += 4;
-            v = ecb_big_endian () ? ecb_rotr32 (v, 8) : ecb_rotl32 (v, 8);
-            *(uint32_t *)line = x; line += 4;
+            uint32_t v = *(uint32_t *)src; src += 4;
+
+            if (ecb_little_endian ())
+              v = ecb_bswap32 (v);
+
+            v = ecb_rotr32 (v, 8);
+
+            if (byte_order_mismatch)
+              v = ecb_bswap32 (v);
+
+            *dst++ = v;
           }
 
       row += rowstride;
+      line += xi.bytes_per_line;
     }
 
   rxvt_img *img = new rxvt_img (s, XRenderFindStandardFormat (dpy, depth == 24 ? PictStandardRGB24 : PictStandardARGB32), 0, 0, width, height);