EXE5 pack ace stuff

.gba

.macro write_pack_byte,value
    .if (value == 0x00)
        ; do nothing
    .elseif (value <= 0x5A)
        add r2, value
    .elseif (value / 2 <= 0x5A)
        add r2, (value / 2)
        add r2, (value / 2)
        .if (value % 2 != 0)
            add r2, (value % 2)
        .endif
    .else
        add r2, (value / 3)
        add r2, (value / 3)
        add r2, (value / 3)
        .if (value % 3 != 0)
            add r2, (value % 3)
        .endif
    .endif
.endmacro

.macro write_pack_word,value
    xor r2, r2
    write_pack_byte ((value >> 0x18) & 0xFF)
    lsl r2, 8
    write_pack_byte ((value >> 0x10) & 0xFF)
    lsl r2, 8
    write_pack_byte ((value >> 0x08) & 0xFF)
    lsl r2, 8
    write_pack_byte ((value >> 0x00) & 0xFF)
.endmacro

.macro write_pack,value
    write_pack_word value
    str r2, [r0, r1]
    add r1, 4
.endmacro

; r0 = pointer to start of pack (+1 b/c we use r0 to jump here)
; r1 = offset added to r0, used for writing
; r2 = byte value written

.create "pack.dat", 0x02002EAC
    ; r0 = 02002EA1 entering the pack
    ; goal is to get 0x203F2FF in r0 (where our playload is going to start)
    write_pack_word 0x0003C45E
    mov r1, r2          ; can't do "add r0, r2" due to one of the opcode bytes being > 0x5A
    add r0, r1          ; r0 = 02002EA1 + 0x3C45E
    mov r1, 1           ; r1 is now the used as the write offset from r0

    ; start writing the actual payload
    write_pack 0x4B02B5FF
    write_pack 0x4D034C02
    write_pack 0xE0074E03
    write_pack 0x04000130
    write_pack 0x00000300
    write_pack 0x0203F32A
    write_pack 0x70286818
    write_pack 0x42203501
    write_pack 0x0000D1F8

    bx r0
.close

; based on http://tasvideos.org/forum/viewtopic.php?p=491739#491739
.create "payload.dat", 0x203F300
    push {r0-r7, lr}
    ldr     r3, =0x04000130
    ldr     r4, =0x300
    ldr     r5, =payload
    ldr     r6, =0x08000391 ; awaitFrame function
    b loop
    .pool

loop:
    mov lr, pc
    bx r6

    ldr     r0, [r3]
    strb    r0, [r5]
    add     r5, 1
    tst     r0, r4
    bne     loop
    ; pop {r0-r7, pc}
payload:
.align 4
.close