From fb2345238378160ae358ce41b3a6fa286f6ee681 Mon Sep 17 00:00:00 2001
From: Irmen de Jong <irmen@razorvine.net>
Date: Fri, 23 Aug 2024 19:00:12 +0200
Subject: [PATCH] optimize monogfx.fill()

---
 compiler/res/prog8lib/cx16/monogfx.p8 | 85 +++++++++++++++++++++------
 examples/test.p8                      | 57 +++++++++++-------
 2 files changed, 103 insertions(+), 39 deletions(-)

diff --git a/compiler/res/prog8lib/cx16/monogfx.p8 b/compiler/res/prog8lib/cx16/monogfx.p8
index 52d675e90..1c66bfbe9 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 +840,70 @@ skip:
                 left = xx
             } until xx>x2
         }
+
+        sub fill_scanline_left() -> bool {
+            ; TODO maybe this could use vera auto decrement, but that would require some clever masking calculations
+            cx16.r9s = xx
+            while xx >= 0 {
+                if pgetset()
+                    break
+                xx--
+            }
+            return xx==cx16.r9s
+        }
+
+        sub fill_scanline_right() {
+             ; TODO maybe this could use vera auto increment, but that would require some clever masking calculations
+             cx16.r9s = xx
+            while xx <= width-1 {
+                if pgetset()
+                    break
+                xx++
+            }
+        }
+
+        sub pgetset() -> bool {
+            ; test and optionally set a pixel
+            word @zp xpos = xx
+            %asm {{
+                lda  p8v_xpos
+                and  #7
+                pha     ; xbits
+            }}
+            xpos /= 8
+            if width==320
+                xpos += yy*(320/8) as uword
+            else
+                xpos += yy*(640/8) as uword
+
+            %asm {{
+                stz  cx16.VERA_CTRL
+                stz  cx16.VERA_ADDR_H
+                lda  p8v_xpos+1
+                sta  cx16.VERA_ADDR_M
+                lda  p8v_xpos
+                sta  cx16.VERA_ADDR_L
+                ply         ; xbits
+                lda  p8s_plot.p8v_maskbits,y
+                and  cx16.VERA_DATA0
+                beq  +
+                lda  #1
++
+                ; cx16.r11L = seed color to check against
+                eor  cx16.r11L
+                beq  +
+                rts
++               ; cx16.r10L = new color to set
+                lda  p8s_plot.p8v_maskbits,y
+                ldx  cx16.r10L
+                beq  +
+                tsb  cx16.VERA_DATA0
+                bra  ++
++               trb  cx16.VERA_DATA0
++               lda  #0
+                rts
+            }}
+        }
     }
 
     sub position(uword @zp xx, uword yy) {
diff --git a/examples/test.p8 b/examples/test.p8
index ca6a7ddd9..af54f1ce6 100644
--- a/examples/test.p8
+++ b/examples/test.p8
@@ -1,25 +1,42 @@
-%option no_sysinit, enable_floats
+%import monogfx
+%import textio
+%import math
+
+%option no_sysinit
 %zeropage basicsafe
 
-main {
-    sub start() {
-        ubyte[] array = [1,2,3]
-        ubyte[3] array2
-        float[] flarray = [1.1, 2.2, 3.3]
-        float[3] flarray2
-        word[] warray = [-2222,42,3333]
-        word[3] warray2
-        str[] names = ["apple", "banana", "tomato"]
-        str[3] names2
 
-        ; 8 array assignments -> 8 arraycopies:
-        array = [8,7,6]
-        array = array2
-        flarray = [99.9, 88.8, 77.7]
-        flarray = flarray2
-        warray = [4444,5555,6666]
-        warray = warray2
-        names = ["x1", "x2", "x3"]
-        names = names2
+main {
+
+    sub start() {
+        monogfx.lores()
+        demofill()
+    }
+
+    sub demofill() {
+        const uword offsetx = 0
+        const uword offsety = 0
+
+        monogfx.circle(offsetx+160, offsety+120, 110, true)
+        monogfx.rect(offsetx+180, offsety+5, 25, 190, true)
+        monogfx.line(offsetx+100, offsety+150, offsetx+240, offsety+10, true)
+        monogfx.line(offsetx+101, offsety+150, offsetx+241, offsety+10, true)
+        monogfx.rect(offsetx+150, offsety+130, 10, 100, true)
+
+        sys.wait(30)
+
+        cbm.SETTIM(0,0,0)
+        monogfx.fill(offsetx+100,offsety+100,true)
+        monogfx.fill(offsetx+100,offsety+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
+
     }
 }