From ef22e5a2fff1cd2a96fb2cf880570fb2a1a517b1 Mon Sep 17 00:00:00 2001
From: hugova <hugova@kth.se>
Date: Sat, 26 Jul 2025 14:50:23 +0200
Subject: [PATCH] use faster memset algoritm.

---
 wip-hugo/routines/memory/memset.s | 118 ++++++++++++++++++++----------
 wip-hugo/source.s                 |   3 +-
 2 files changed, 80 insertions(+), 41 deletions(-)
 mode change 100755 => 100644 wip-hugo/routines/memory/memset.s

diff --git a/wip-hugo/routines/memory/memset.s b/wip-hugo/routines/memory/memset.s
old mode 100755
new mode 100644
index 55700e9..37371ad
--- a/wip-hugo/routines/memory/memset.s
+++ b/wip-hugo/routines/memory/memset.s
@@ -1,40 +1,78 @@
-;;; -*- Mode: asm; indent-tabs-mode: t; tab-width: 8 -*-
-
-;; Sets memory in 'A'-registry to all addresses from 'A_start' until 'A_start' + 'length'
-;; Modifies A, X and A_start
-.proc memset
-        .include "mem.inc"
-
-;; big_set sets the memory in $ff chunks.
-;; skipp if length >= $ff
-LDX length +1
-BNE big_set
-JMP small_set
-
-big_set: ;sets $ff of memory
-        ;; Y value do not matter, will go through all anyway!
-        .repeat $ff
-        STA (A_start), Y
-        DEY
-        .endrepeat
-        STA (A_start), Y ; dont forget Y =0
-big_set_end:
-        ;;set all hole $ff memory chunks!
-        INC A_start + 1
-        DEX ;; length +1 --
-        BEQ small_set
-        JMP big_set
-
-
-
-;;sets the rest of the memory
-;; note that this can use code above (smc) or the same method. may implement later.
-small_set:
-       LDY length
-small_set_loop:
-        STA (A_start), Y
-        DEY
-        BNE small_set_loop
-        STA (A_start), Y
-        RTS
-.endproc
+;;; -*- Mode: asm; indent-tabs-mode: t; tab-width: 8 -*-
+
+;; Sets memory in 'A'-registry to all addresses from 'A_start' until 'A_start' + 'length'
+;; Modifies A, X and A_start
+.proc memset
+        .include "mem.inc"
+
+;; big_set sets the memory in $ff chunks.
+;; skipp if length >= $ff
+LDX length +1
+BNE big_set
+JMP small_set
+
+big_set: ;sets $ff of memory
+        ;; Y value do not matter, will go through all anyway!
+        .repeat $ff
+        STA (A_start), Y
+        DEY
+        .endrepeat
+        STA (A_start), Y ; dont forget Y =0
+big_set_end:
+        ;;set all hole $ff memory chunks!
+        INC A_start + 1
+        DEX ;; length +1 --
+        BEQ small_set
+        JMP big_set
+
+;; Note that cpu cykels total: cy_tot = 66 to 69
+;; But we skipp a BNE (cy = 2*) * [length (mod 255)]
+;; The BNE case has an avrige of 2*255/2 = 255 so this is faster (on avrige.)
+
+small_set:
+        STA data_to_write               ; cy = 3
+        LDA length                      ; cy = 3
+        STA length_copy                 ; cy = 3
+
+        ;; calculate rts-position
+        LDX #$00                        ; cy = 2
+        STX length + 1                  ; cy = 3
+        ;; 3 bytes = STA DEY NOP = seting 1 byte of memory.
+        ;; So we need to calculate: length*3
+        Mult_16 A, length + 1           ; cy = 7
+        ; A=  length
+        ADC length_copy                 ; cy = 3
+        TAY
+        LDA length + 1                  ; cy = 3
+        ADC #$00                        ; cy = 2
+        STA length + 1                  ; cy = 3
+
+        ;; Now RTS_pointer + Y = length*3 + big_set_label
+        LDA #<big_set                   ; cy = 2
+        STA RTS_pointer                 ; cy = 3
+        LDA #>big_set                   ; cy = 2
+        ADC length + 1                  ; cy = 3
+        STA RTS_pointer + 1             ; cy = 3
+
+        ;; read data we will change to RTS
+        STY Y_copy                      ; cy = 3
+        LDA (RTS_pointer), Y            ; cy = 5*
+        TAX                             ; cy = 2
+
+        ;; set RTS in big_set
+        LDA #$60                        ; cy = 2
+        STA (RTS_pointer), Y            ; cy = 5*
+
+        ;; JSR to modified big_set
+        LDY length_copy                 ; cy = 3
+        DEY ; because we want to count to Y=0 :)
+        LDA data_to_write               ; cy = 3
+        JSR big_set                     ; cy = 6
+
+        ;; revert changes
+        LDY Y_copy                      ; cy = 3
+        TXA                             ; cy = 2
+        STA (RTS_pointer), Y            ; cy = 5*
+
+        RTS
+.endproc
diff --git a/wip-hugo/source.s b/wip-hugo/source.s
index 9c0bfd3..59c0c30 100644
--- a/wip-hugo/source.s
+++ b/wip-hugo/source.s
@@ -33,7 +33,8 @@ JMP exit
 .include "routines/pixel/pixel_draw.s"
 .include "routines/pixel/pixel_calc.s"
 .include "routines/text/char_draw.s"
-.include "routines/memory/memset_alt.s"
+.include "routines/memory/memset.s"
+;.include "routines/memory/clear_screen.s"
 .include "routines/memory/memcpy.s"
 .include "routines/arithmatic/mult.s"
 .include "routines/arithmatic/div.s"