;;; -*- 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
        NOP
        NOP
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