diff --git a/compiler/res/prog8lib/cx16/monogfx.p8 b/compiler/res/prog8lib/cx16/monogfx.p8
index 52d675e90..c49639cbc 100644
--- a/compiler/res/prog8lib/cx16/monogfx.p8
+++ b/compiler/res/prog8lib/cx16/monogfx.p8
@@ -819,31 +819,14 @@ invert:
         while cx16.r12L!=0 {
             pop_stack()
             xx = x1
-            while xx >= 0 {
-                if pget(xx as uword, yy as uword) as ubyte != cx16.r11L
-                    break
-                xx--
-            }
-            if x1!=xx
-                horizontal_line(xx as uword+1, yy as uword, x1-xx as uword, cx16.r10L as bool)
-            else
-                goto skip
-
+            if fill_scanline_left() goto skip
             left = xx + 1
             if left < x1
                 push_stack(left, x1 - 1, yy, -dy)
             xx = x1 + 1
 
             do {
-                cx16.r9s = xx
-                while xx <= width-1 {
-                    if pget(xx as uword, yy as uword) as ubyte != cx16.r11L
-                        break
-                    xx++
-                }
-                if cx16.r9s!=xx
-                    horizontal_line(cx16.r9, yy as uword, xx-cx16.r9s as uword, cx16.r10L as bool)
-
+                fill_scanline_right()
                 push_stack(left, xx - 1, yy, dy)
                 if xx > x2 + 1
                     push_stack(x2 + 1, xx - 1, yy, -dy)
@@ -857,6 +840,33 @@ skip:
                 left = xx
             } until xx>x2
         }
+
+        sub fill_scanline_left() -> bool {
+            ; TODO optimize this to use vera auto-decrements, but requires masking etc because of 8 pixels per byte...
+            cx16.r9s = xx
+            while xx >= 0 {
+                if pget(xx as uword, yy as uword) as ubyte != cx16.r11L
+                    break
+                xx--
+            }
+            if xx!=cx16.r9s {
+                horizontal_line(xx+1 as uword, yy as uword, cx16.r9s-xx as uword, cx16.r10L as bool)
+                return false
+            }
+            return true
+        }
+
+        sub fill_scanline_right() {
+            ; TODO optimize this to use vera auto-increments, but requires masking etc because of 8 pixels per byte...
+            cx16.r9s = xx
+            while xx <= width-1 {
+                if pget(xx as uword, yy as uword) as ubyte != cx16.r11L
+                    break
+                xx++
+            }
+            if xx!=cx16.r9s
+                horizontal_line(cx16.r9, yy as uword, xx-cx16.r9s as uword, cx16.r10L as bool)
+        }
     }
 
     sub position(uword @zp xx, uword yy) {
diff --git a/examples/test.p8 b/examples/test.p8
index 10a85aec4..9ffd45632 100644
--- a/examples/test.p8
+++ b/examples/test.p8
@@ -1,37 +1,39 @@
-%import palette
+%import monogfx
 %import textio
+%import math
+
 %option no_sysinit
+%zeropage basicsafe
+
 
 main {
+
     sub start() {
-        repeat 4 {
-            for cx16.r0L in 0 to 15 {
-                txt.color2(cx16.r0L, cx16.r0L)
-                txt.spc()
-                txt.spc()
-                txt.spc()
-                txt.spc()
-            }
-            txt.nl()
-        }
-        bool changed
-        uword[] colors = [
-            $f00, $800, $200, $000,
-            $f0f, $80f, $20f, $00f
-        ]
-        do {
-            sys.waitvsync()
-            sys.waitvsync()
-            changed = palette.fade_step_colors(0, 8, colors)
-        } until not changed
-
-        sys.wait(60)
-        changed = false
-        do {
-            sys.waitvsync()
-            sys.waitvsync()
-            changed = palette.fade_step_multi(0, 8, $fff)
-        } until not changed
-        sys.wait(60)
+        monogfx.lores()
+        demofill()
+    }
+
+    sub demofill() {
+        monogfx.circle(160, 120, 110, true)
+        monogfx.rect(180, 5, 25, 190, true)
+        monogfx.line(100, 150, 240, 10, true)
+        monogfx.line(101, 150, 241, 10, true)
+        monogfx.rect(150, 130, 10, 100, true)
+
+        sys.wait(30)
+
+        cbm.SETTIM(0,0,0)
+        monogfx.fill(100,100,true)
+        monogfx.fill(100,100,false)
+        uword duration = cbm.RDTIM16()
+        sys.wait(30)
+
+        monogfx.textmode()
+        txt.nl()
+        txt.print_uw(duration)
+        txt.print(" jiffies\n")
+
+        ; before optimizations: ~166 jiffies
+
     }
 }