add alternate version of memset (probably slower (only for testing))

wip-hugo/routines/circle/circle.inc

@@ -8,14 +8,14 @@
         t2    = $E1
         X_math = radius
         Y_math = $E2
-        ;; E9 - EA is used by pixel.inc
-        byte_to_paint_qaa = $E8
-        byte_to_paint_qcb = $E3
-        byte_to_paint_qca = $E4
-        byte_to_paint_qdb = $E5
+        byte_to_paint_qaa = $E3
+        byte_to_paint_qcb = $E4
+        byte_to_paint_qca = $E5
+        byte_to_paint_qdb = $E6
 
         btp_mem_pos_qaa = btp_mem_pos
-        btp_mem_pos_qcb = $E6 ; 16bit value (uses E7)
+        btp_mem_pos_qcb = $E7 ; 16bit value (uses E8)
+                              ; E9 - EA is used by pixel.inc
         btp_mem_pos_qdb = $EB ; 16bit value (uses EC)
         btp_mem_pos_qda = $ED ; 16bit value (uses EE)
         btp_mem_pos_qab = $D0 ; 16bit value (uses D1)

wip-hugo/routines/memory/mem.inc

@@ -4,3 +4,8 @@
         A_start = $D0      ; 16-bit value (uses D1)
         B_start = $D2      ; 16-bit value (uses D3)
         length = $D4       ; 16-bit value (uses D5), this is the number of bytes
+
+        RTS_pointer = $D6  ; 16-bit value (uses D7)
+        data_to_write = $D8
+        length_copy = $d9  ; 16-bit vale (uses da)
+        instruction_backup = $db

wip-hugo/routines/memory/memset.s

@@ -18,6 +18,8 @@ big_set: ;sets $ff of memory
         DEY
         .endrepeat
         STA (A_start), Y ; dont forget Y =0
+        NOP
+        NOP
 big_set_end:
         ;;set all hole $ff memory chunks!
         INC A_start + 1
@@ -26,15 +28,46 @@ big_set_end:
         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
+        STA data_to_write
+        LDA length
+        STA length_copy
+
+        ;; calculate rts-position
+        LDA #$00
+        STA length + 1
+        ;; 4 bytes = STA DEY NOP = seting 1 byte of memory.
+        ;; So we need to calculate: length*4
+        Mult_16 length, length + 1
+        Add_16 length, length + 1, length_copy, #$00
+        ;; Now RTS_pointer = length*4 + big_set_label
+        CLC
+        LDA #<big_set
+        ADC length
+        STA RTS_pointer
+        LDA #>big_set
+        ADC length + 1
+        STA RTS_pointer + 1
+
+        ;; read data we will change to RTS
+        LDY #$00
+        LDA (RTS_pointer), Y
+        STA instruction_backup
+
+        ;; set RTS in big_set
+        LDA #$60
+        STA (RTS_pointer), Y
+
+        ;; JSR to modified big_set
+        LDY length_copy
+        DEY ; because we want to count to Y=0 :)
+        LDA data_to_write
+        JSR big_set
+
+        ;; revert changes
+        LDY #$00
+        LDA instruction_backup
+        STA (RTS_pointer), Y
+
         RTS
 .endproc

wip-hugo/routines/memory/memset_alt.s

@@ -0,0 +1,71 @@
+;;; -*- 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
+
+
+small_set:
+        STA data_to_write
+        LDA length
+        STA length_copy
+
+        ;; calculate rts-position
+        LDA #$00
+        STA length + 1
+        ;; 4 bytes = STA DEY NOP = seting 1 byte of memory.
+        ;; So we need to calculate: length*4
+        Mult_16 length, length + 1
+        Add_16 length, length + 1, length_copy, #$00
+        ;; Now RTS_pointer = length*4 + big_set_label
+        CLC
+        LDA #<big_set
+        ADC length
+        STA RTS_pointer
+        LDA #>big_set
+        ADC length + 1
+        STA RTS_pointer + 1
+
+        ;; read data we will change to RTS
+        LDY #$00
+        LDA (RTS_pointer), Y
+        STA instruction_backup
+
+        ;; set RTS in big_set
+        LDA #$60
+        STA (RTS_pointer), Y
+
+        ;; JSR to modified big_set
+        LDY length_copy
+        DEY ; because we want to count to Y=0 :)
+        LDA data_to_write
+        JSR big_set
+
+        ;; revert changes
+        LDY #$00
+        LDA instruction_backup
+        STA (RTS_pointer), Y
+
+        RTS
+.endproc