15,233 bytes of code and 17,535 of data. 7,509 instructions spread over 985 named blocks, and not one below 10 % commented.
INIT (0x407E) does no more than set the stage: it switches the whole cartridge into pages 1 and 2 with ENASLT, writes a jp cada_cuadro into the H.KEYI hook at 0xFD9A, clears 0xE000 to 0xE7FF, builds the screen and drops into a jr $. The whole game happens inside the interrupt.
cada_cuadro (0x402E) does, in this order: read the VDP to drop the interrupt request, sound first — so its delay does not show if the frame runs long — the reentry latch, the controls, the frame, and one more VDP read on the way out.
Almost all the flow goes through 0x406C:
pop hl ; the return address IS the table
add a,a
call suma_a_a_hl
ld e,(hl) / inc hl / ld d,(hl) / ex de,hl / jp (hl)So the table sits right behind the call. There are fifteen such places, and tools/tablas_despacho.py works out each table's size with the lowest-entry rule. Three would not yield: their first entry jumps backwards to code nothing else calls, and the rule has nothing to stand on. They were closed by hand on the first entry pointing forwards, which is always the byte right after the table.
There is a second door, 0x4065, which takes the index from (ix+0) and wants the table already in HL. Only one place uses it, 0x75F0, and that is the two-level dispatch: 0x75F3 is a table of sixteen subtables indexed by (ix+1) & 0x0F — the enemy type — and 0x4065 indexes the subtable with (ix+0), which is its state.
(0xE000) is the scene and (0xE001) the subscene; the table at 0x40EA has ten entries:
0 intro (the KONAMI banner coming down)
1 title
2 demo
3 countdown and start of the game
4 the game <- the scene the Game Master watches
5 stage cleared
6 a life lost
7 GAME OVER
8 next stage
9 the endingIn scenes 0, 1 and 2, haz_el_cuadro pushes 0x4262 onto the stack before dispatching, so on return from the scene that runs: look at whether anything was pressed and, if we are on the title and it is a fire button, start the game.
There are two formats, and it pays not to mix them up.
The block one (0x4417), for patterns, colour and tiles:
n with bit 7 set n bytes as they come
n without bit 7 the next byte, repeated n times
0x80 change of VRAM address (two bytes)
0x00 endEvery byte also goes through lee_y_traduce (0x443B), which applies the colour bank in (0xE661).
The label one (0x43F9): a word with the VRAM address, then the tile codes, 0xFE to jump somewhere else and 0xFF to finish. The codes are not ASCII: they are tile numbers of the font, and the alphabet is read by drawing the 35 glyphs 0x45C9 decompresses.
It is The Goonies' (RC-734), and not by hearsay: tools/porta_nombres.py finds its routines here by normalised instruction signature, with runs of up to 150 bytes in a row.
Three fourteen-byte voices at 0xE010, 0xE01E and 0xE02C. A piece is asked for by number, and that number is also its priority: it only gets in if it outweighs whatever is already playing. The table at 0x4B38 is 63 voice pointers, not piece pointers: each piece takes one, two or three in a row.
The twelve semitones are at 0x4B2C, and dropping an octave means doubling the period: the add hl,hl at 0x4B0F.
tools/comun_normalizado.py gives 21 runs and 996 normalised bytes in common with The Goonies, that is 6.8 % of this cartridge. What is shared is the framework — the dispatcher, the 16-bit adds, the decompressors, the sprite mirror, the controller reader — and the sound engine. The other 13.7 KB of code are its own.