--- rxvt-unicode/src/rxvtimg.C	2012/07/02 02:01:41	1.100
+++ rxvt-unicode/src/rxvtimg.C	2012/07/14 08:26:56	1.103
@@ -172,9 +172,9 @@
     }
 
     ecb_noinline
-    void mask (bool rgb = true, int x = 1, int y = 1)
+    void mask (bool rgb = true, int w = 1, int h = 1)
     {
-      Pixmap pixmap = XCreatePixmap (dpy, srcimg->pm, x, y, rgb ? 32 : 8);
+      Pixmap pixmap = XCreatePixmap (dpy, srcimg->pm, w, h, rgb ? 32 : 8);
 
       XRenderPictFormat *format = XRenderFindStandardFormat (dpy, rgb ? PictStandardARGB32 : PictStandardA8);
       XRenderPictureAttributes pa;
@@ -351,35 +351,28 @@
       unsigned char *src = row;
       uint32_t      *dst = (uint32_t *)line;
 
-      if (!pb_has_alpha)
-        for (int x = 0; x < width; x++)
-          {
-            uint8_t r = *src++;
-            uint8_t g = *src++;
-            uint8_t b = *src++;
-
-            uint32_t v = (255 << 24) | (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 *)src; src += 4;
-
-            if (ecb_big_endian ())
-              v = ecb_bswap32 (v);
+      for (int x = 0; x < width; x++)
+        {
+          uint8_t r = *src++;
+          uint8_t g = *src++;
+          uint8_t b = *src++;
+          uint8_t a = *src;
+
+          // this is done so it can be jump-free, but newer gcc's clone inner the loop
+          a = pb_has_alpha ? a : 255;
+          src += pb_has_alpha;
 
-            v = ecb_rotl32 (v, 8); // abgr to bgra
+          r = (r * a + 127) / 255;
+          g = (g * a + 127) / 255;
+          b = (b * a + 127) / 255;
 
-            if (!byte_order_mismatch)
-              v = ecb_bswap32 (v);
+          uint32_t v = (a << 24) | (r << 16) | (g << 8) | b;
 
-            *dst++ = v;
-          }
+          if (ecb_big_endian () ? !byte_order_mismatch : byte_order_mismatch)
+            v = ecb_bswap32 (v);
+
+          *dst++ = v;
+        }
 
       row += rowstride;
       line += xi.bytes_per_line;
@@ -596,6 +589,64 @@
   return img;
 }
 
+rxvt_img *
+rxvt_img::muladd (nv mul, nv add)
+{
+  // STEP 1: double the image width, fill all odd columns with white (==1)
+
+  composer cc (this, new rxvt_img (s, format, 0, 0, w * 2, h, repeat));
+
+  // why the hell does XRenderSetPictureTransform want a writable matrix :(
+  // that keeps us from just static const'ing this matrix.
+  XTransform h_double = {
+    0x8000, 0,     0,
+    0, 0x1000,     0,
+    0,     0, 0x1000
+  };
+
+  XRenderSetPictureFilter (cc.dpy, cc.src, "nearest", 0, 0);
+  XRenderSetPictureTransform (cc.dpy, cc.src, &h_double);
+  XRenderComposite (cc.dpy, PictOpSrc, cc.src, None, cc.dst, 0, 0, 0, 0, 0, 0, w * 2, h);
+
+  cc.mask (false, 2, 1);
+
+  static const XRenderColor c0 = {     0,     0,     0,     0 };
+  XRenderFillRectangle (cc.dpy, PictOpSrc, cc.msk, &c0, 0, 0, 1, 1);
+  static const XRenderColor c1 = { 65535, 65535, 65535, 65535 };
+  XRenderFillRectangle (cc.dpy, PictOpSrc, cc.msk, &c1, 1, 0, 1, 1);
+
+  Picture white = XRenderCreateSolidFill (cc.dpy, &c1);
+
+  XRenderComposite (cc.dpy, PictOpOver, white, cc.msk, cc.dst, 0, 0, 0, 0, 0, 0, w * 2, h);
+
+  XRenderFreePicture (cc.dpy, white);
+
+  // STEP 2: convolve the image with a 3x1 filter
+  // a 2x1 filter would obviously suffice, but given the total lack of specification
+  // for xrender, I expect different xrender implementations to randomly diverge.
+  // we also halve the image, and hope for the best (again, for lack of specs).
+  composer cc2 (cc.dstimg);
+
+  XFixed kernel [] = {
+    XDoubleToFixed (3), XDoubleToFixed (1),
+    XDoubleToFixed (0), XDoubleToFixed (mul), XDoubleToFixed (add)
+  };
+
+  XTransform h_halve = {
+    0x2000, 0,      0,
+    0, 0x1000,      0,
+    0,      0, 0x1000
+  };
+
+  XRenderSetPictureFilter (cc.dpy, cc2.src, "nearest", 0, 0);
+  XRenderSetPictureTransform (cc.dpy, cc2.src, &h_halve);
+  XRenderSetPictureFilter (cc.dpy, cc2.src, FilterConvolution, kernel, ecb_array_length (kernel));
+
+  XRenderComposite (cc.dpy, PictOpSrc, cc2.src, None, cc2.dst, 0, 0, 0, 0, 0, 0, w * 2, h);
+
+  return cc2;
+}
+
 ecb_noinline static void
 extract (int32_t cl0, int32_t cl1, int32_t &c, unsigned short &xc)
 {
@@ -709,8 +760,6 @@
   if (x == 0 && y == 0 && w == ref->w && h == ref->h)
     return clone ();
 
-  Display *dpy = s->dpy;
-
   // add an alpha channel if...
   bool alpha = !format->direct.alphaMask // pixmap has none yet
                && (x || y)               // we need one because of non-zero offset
@@ -719,7 +768,7 @@
   composer cc (this, new rxvt_img (s, alpha ? find_alpha_format_for (s->dpy, format) : format,
                                    0, 0, w, h, repeat));
 
-  if (alpha)
+  if (repeat == RepeatNone)
     {
       XRenderColor rc = { 0, 0, 0, 0 };
       XRenderFillRectangle (cc.dpy, PictOpSrc, cc.dst, &rc, 0, 0, w, h);//TODO: split into four fillrectangles