.long fakestack, _start
.global _start
_start:
  move.w #0x1, d0 | Find the ROM card
  bra.b find_first_card
romfindret:
  move.l a0, a2 | Save the ROM card IO base in a6 for later
  move.l #0x10000, a7 | Set up the stack at the end of the ROM card's RAM
  bsr.w find_largest_ram | Find the largest RAM card and put the IO base in a3
  move.l a0, a3
  move.l d0, d2
  move.l #0x8100, a1
  bsr.w find_all_ram_cards
  move.l #0x8100, a0
  bsr.w sort_ram_cards
  move.w #0x4, d0 | Find a storage card and put the IO base in a4
  bsr.b find_first_card
  move.l a0, a4
  | Transfer the bootsector load code to the ROM's
  | built in RAM at the start of it's IO space
  move.w #(ramcode_end - ramcode), d0 | Put the length of the ramcode in d0
  move.w #ramcode, a0 | Put the address of the ramcode in a0
  move.l a2, a1 | Put the start of the ROM's IO space RAM in a0
  ramcode_loop:
  move.b (a0)+, (a1)+ | Transfer a byte of ramcode to the ROM's IO space RAM
  dbra d0, ramcode_loop | Loop back if there is more to transfer
  jmp (a2) | Jump to the ramcode

ramcode:
  move.b #0x0, (0xF3, a2) | Disable the ROM
  move.l #0x1, (a3) | Enable the RAM at base 0x0
  cmpi.l #0x1000000, d2
  bcs.b sp_ok
  move.l #0xff0000, d2
  sp_ok:
  move.l d2, a7
  | Load sector 0 to 0x0
  move.l #0x0, (0x0, a4) | Set the sector number to 0
  move.l #0x1, (0x4, a4) | Set the sector count to 1
  move.l #0x0, (0xC, a4) | Set the destination address to 0x0
  move.w #0x1, (0x8, a4) | Send a DMA read command
  jmp (0x0).W | Jump to the loaded sector
ramcode_end:

nop | Padding to make sure ramcode_end and find_first_card are different

| Finds the first card with the type in d0.w, and returns it's IO base address in a0, or 0 if not found
find_first_card:
  move.l #0xff0000, a0 | a0 holds the address of the current card
ffc_loop:
  lea (0x100,a0), a0 | adda.l #$100,a0 ; Move to the next card
  move.w (0xfe, a0), d1 | Load the type of the card into d1
  beq.b ffc_done | If the type is 0 (empty slot), we have scanned all cards, so exit the loop
  cmp.w d0, d1 | If the card is the type we want, return with the address in a0
  beq.b ffc_done
  bra.b ffc_loop | Loop back and check the next card
ffc_done:
  rts

| Finds the largest RAM card, and returns it's IO base address in a0 and size in d0, or 0 if not found
find_largest_ram:
  move.l #0x0, d0 | d0 holds the size of the largest RAM card found
  move.w #0x0, a0 | a0 holds the address of the largest RAM card found
  move.l #0xff0000, a1 | a1 holds the address of the current card
flr_loop:
  lea (0x100,a1), a1 | adda.l #$100,a0 ; Move to the next card
  move.w (0xfe, a1), d1 | Load the type of the card into d1
  beq.b flr_done | If the type is 0 (empty slot), we have scanned all cards, so exit the loop
  cmp.w #0x2, d1 | If the card isn't a RAM card, skip it
  bne.b flr_loop
  move.l (0x4, a1), d1 | Load the card's size into d1
  cmp.l d0, d1 | If the current size is less than the largest size found, go back to the start of the loop
  bls.b flr_loop
  move.l d1, d0 | Store the size and address of the new largest card in d0 and a0
  move.l a1, a0
  bra.b flr_loop | Loop back and check the next card
flr_done:
  rts
fakestack: 
  .long romfindret

| 1KB buffer in a1
| Returns list length in d0
find_all_ram_cards:
  move.b #0, d0 | d0 holds the number of RAM cards found.
  move.l #0xff0000, a0 | a0 holds the address of the current card.
farc_loop:
  lea (0x100, a0), a0 | adda.l #$100,a0 ; Move to the next card
  move.w (0xfe, a0), d1 | Load the type of the card into d1
  beq.b farc_done | If the type is 0 (empty slot), we have scanned all cards, so exit the loop
  cmp.w #0x2, d1 | If the card isn't a RAM card, skip it
  bne.b farc_loop 
  move.l a0, (a1)+ | Write the IO base address into the buffer and advance the buffer pointer
  addq.b #1, d0 | Increment the count of found RAM cards
  bra.b farc_loop | Loop back and check the next card
farc_done:
  move.l #0, (a1) | Write a null terminator on the end of the list
  rts

| optimized bubble sort to sort RAM cards by size
| Buffer in a0
| Length in d0
sort_ram_cards:
movem.l d2-d4/a2, -(a7) | Save calle-preserved registers
move.b #0, d1 | d1 = newlen
move.b #1, d2 | d2 = i
move.l a0, a1 | a1 holds the pointer to the current pair
src_inner_loop:
cmp.b d0, d2 | if i >= length, exit the inner loop
bge.b src_inner_loop_done
move.l (a1), a2 | Read the first card's size
move.l (0x4, a2), d3
move.l (4, a1), a2 | Read the second card's size
move.l (0x4, a2), d4
cmp.l d3, d4 | if d4 <= d3, branch to src_pair_sorted
bls.b src_pair_sorted
move.l (a1), d3 | Read the first value
move.l (0x4, a1), d4 | Read the second value
move.l d3, (0x4, a1) | Write the first value where the second value was
move.l d4, (a1)+ | Write the second value where the first value was and increment a1 for the next pair
move.b d2, d1 | i = newlem
src_pair_sorted:
addq.b #1, d2 | i++
bra.b src_inner_loop
src_inner_loop_done:
move.b d1, d0 | len = newlen
cmp #1, d0 | if n > 1, branch to outer loop
bgt.b sort_ram_cards
movem.l (a7)+, d2-d4/a2 | Restore calle-preserved registers
rts