processor 6502

srcPointer = $b5
dstPointer = $b7
dstPointerEnd = $b9
fake_current_byte = $126
count = $128

    org $a000

    ; autostart cartridge magic number
    WORD coldstart
    WORD coldstart
    BYTE #$41, #$30, #$c3, #$c2, #$cd

coldstart SUBROUTINE
    cld
    sei

    jsr $fd8d   ; initialize and test RAM
    jsr $fd52   ; restore default I/O vectors
    jsr $fdf9   ; initialize I/O registers
    jsr $e518   ; initialize hardware
    cli         ; enable interrupts

    lda $0288
    sta $0294

    jsr $e45b   ; initialize BASIC vector table
    jsr $e3a4   ; initialize BASIC RAM locations
    jsr $e404   ; print startup message and initialize memory pointers
    ldx #$fb    ; value for stack start
    txs         ; set stack pointer
    lda #$76    ; set "READY." pointer
    ldy #$c3
    jsr $cb1e   ; print null terminated string
    lda #$80    ; set for control messages only
    jsr $ff90   ; control kernal messages

.oracopio:
    jsr copycode
.almost_ready:

    ; set current BASIC line number
    lda #$ff
    sta $3a
    lda #0
    sta $3b

    ; set start of variables
    lda #<($1001 + demo_end - demo_start)
    sta $2d
    lda #>($1001 + demo_end - demo_start)
    sta $2e

    jsr clear_color_ram

    ; print informative message for user
    lda #<message
    ldy #>message
    jsr $cb1e       ; print null terminated string

    ; install fake chrin to hijack custom commands
    sei
    lda #<fake_chrin
    sta $324
    lda #>fake_chrin
    sta $325
    cli

    jmp $c483 ; BASIC

    ; custom hijacked command line
fake_input_line:
    BYTE "RUN",#$0d,#0,#0,#0

fake_chrin SUBROUTINE
    ldy fake_current_byte
    lda fake_input_line,y
    iny
    sty fake_current_byte
    cpy #5
    bne .end_fake_chrin
    dey
    sty fake_current_byte

.end_fake_chrin:
    clc
    rts


clear_color_ram SUBROUTINE
    lda #$02
    ldx #$00
.loop:
    sta $9600,x
    sta $9700,x
    dex
    bne .loop
    rts

; This routine does a lot of unnecessary things,
; but I added a lot of garbage things to debug it,
; because it did not work as intended.
; In the end, the problem was elsewhere, as usual.
copycode SUBROUTINE
    sei
    lda #0
    sta count
    sta count + 1

    lda #<demo_start
    sta srcPointer
    lda #>demo_start
    sta srcPointer + 1

    lda #$01
    sta dstPointer
    lda #$10
    sta dstPointer + 1

    ldy #0
copyloop:
    lda (srcPointer),y
    sta (dstPointer),y
    inc srcPointer
    bne .noinc1
    inc srcPointer + 1
.noinc1:
    inc dstPointer
    bne .noinc2
    inc dstPointer + 1
.noinc2:

    inc count
    bne .noinc3
    inc count + 1
.noinc3:

    ; lenght of program to copy
    lda count
    cmp #<DEMO_LEN
    bne copyloop
    lda count + 1
    cmp #>DEMO_LEN
    bne copyloop

    cli
.finito:
    rts

message:
    BYTE #$0d,#$0d,"PLEASE WAIT 20 SEC",#$0d,"I'M JUST",#$0d,"AN OLD COMPUTER",#$0d,#0

; https://www.commodore.ca/manuals/funet/cbm/vic20/demos.basic/unexpanded/Stillenacht.prg
demo_start:
    .incbin "build/demo.raw"
demo_end:
DEMO_LEN SET demo_end - demo_start
    ECHO DEMO_LEN

. = $bffc
    jmp coldstart

. = $bfff
    BYTE #$ff