l65/nes.l65

-- set cpu to 6502
cpu = require "6502"
setmetatable(_ENV, cpu)

nes = {
    OAM         = 0x200, -- 0x100 bytes
    RAM         = 0x300, -- 0x500 bytes + ZP 0x100 bytes + Stack 0x100 bytes + OAM 0x100 bytes = 0x800 bytes

    -- 2C02 / 2C07 PPU
    PPUCTRL     = 0x2000,
    PPUMASK     = 0x2001,
    PPUSTAT     = 0x2002,
    OAMADDR     = 0x2003,
    OAMDATA     = 0x2004,
    BGSCROL     = 0x2005,
    PPUADDR     = 0x2006,
    PPUDATA     = 0x2007,

    -- 2A03 / 2A07 CPU+APU
    SQ1VOL      = 0x4000,
    SQ1SWEEP    = 0x4001,
    SQ1LO       = 0x4002,
    SQ1HI       = 0x4003,
    SQ2VOL      = 0x4004,
    SQ2SWEEP    = 0x4005,
    SQ2LO       = 0x4006,
    SQ2HI       = 0x4007,
    TRILINEAR   = 0x4008,
    TRILO       = 0x400A,
    TRIHI       = 0x400B,
    NOISEVOL    = 0x400C,
    NOISELO     = 0x400E,
    NOISEHI     = 0x400F,
    DMCFREQ     = 0x4010,
    DMCRAW      = 0x4011,
    DMCSTART    = 0x4012,
    DMCLEN      = 0x4013,
    OAMDMA      = 0x4014,
    SNDCNT      = 0x4015,
    SPECIO1     = 0x4016,
    SPECIO2     = 0x4017,

    SRAM        = 0x6000, -- 0x2000 bytes
    ROM         = 0x8000, -- 0x8000 bytes

    -- PPU Memory declarations
    CHAR0       = 0x0000, -- 0x1000 bytes
    CHAR1       = 0x1000, -- 0x1000 bytes
    SCREEN0     = 0x2000, -- 0x400 bytes
    SCREEN1     = 0x2400, -- 0x400 bytes
    SCREEN2     = 0x2800, -- 0x400 bytes
    SCREEN3     = 0x2C00, -- 0x400 bytes
    BGPAL       = 0x3F00, -- 0x10 bytes
    OBJPAL      = 0x3F10, -- 0x10 bytes
}
do
    local symbols = cpu.symbols
    for k,v in pairs(nes) do symbols[k] = v end
end

-- add some symbol file formats for NES debuggers
cpu.getsym_as.mesen = function() -- .mlb
    local ins,fmt = table.insert,string.format
    local s = getsym(function(a,l)
        if a >= 0x10000 then return end
        local prefix = {}
        if a < 0x2000 then prefix[1]='R'
        elseif a >= 0x6000 and a < 0x8000 then prefix[1]='S' prefix[2]='W' a=a-0x6000
        elseif a >= 0x8000 then prefix[1]='P' a=a-0x8000
        else prefix[1]='G' end
        local s = {}
        for _,p in ipairs(prefix) do ins(s, fmt("%s:%04x:%s", p, a, l)) end
        return s
    end)
    return table.concat(s, '\n')
end
cpu.getsym_as.fceux = function(filename) -- .nl, multiple files
    local ins,fmt = table.insert,string.format
    local ram,rom = {},{}
    local s = getsym(function(a,l)
        local s = fmt("$%04x#%s#", a, l)
        if a < 0x8000 then ins(ram, s)
        elseif a < 0x10000 then
            local a_org = symbolsorg[l] or a
            local bank = math.floor((a_org - 0x8000) / 0x4000)
            if not rom[bank] then rom[bank] = {} end
            ins(rom[bank], s)
        end
    end)
    local fn = filename
    if not fn:find('%.') then fn = fn .. '.nes' end
    local fni = fn .. '.ram.nl'
    local f = assert(io.open(fni, "wb"), "failed to open " .. fni .. " for writing")
    f:write(table.concat(ram, '\n')) f:close()
    for k,v in pairs(rom) do
        fni = fn .. '.' .. k .. '.nl'
        f = assert(io.open(fni, "wb"), "failed to open " .. fni .. " for writing")
        f:write(table.concat(v, '\n')) f:close()
    end
end

mappers = {}

vblank_waitbegin = function()
    local l=label() bit PPUSTAT bpl l
end
vblank_waitend = function()
    local l=label() bit PPUSTAT bmi l
end

ppu_addr = function(addr)
    lda #addr>>8 sta PPUADDR
    if addr&0xff ~= addr>>8 then lda #addr&0xff end
    sta PPUADDR
end

oam_bytes = function(t)
    return {
        t.y - 1,
        t.tile,
        (t.palette or 0)&3 | ((t.behind or t.priority==1) and 0x20 or 0) | (t.flipx and 0x40 or 0) | (t.flipy and 0x80 or 0),
        t.x
    }
end
oam_set = function(t)
    local b = oam_bytes(t)
    lda #t[1]*4 sta OAMADDR
    for _,v in ipairs(b) do lda #v sta OAMDATA end
end

oamcache = OAM -- change it to set other location
oamcache_clear = function()
    local oam = oamcache
    ldx #0 lda #0xff
    local l=label() sta oam,x inx inx inx inx bne l
end
oamcache_flush = function()
    local oam = oamcache
    lda #0 sta OAMADDR lda #oam>>8 sta OAMDMA
end
oamcache_set = function(t)
    local oam = oamcache
    local b = oam_bytes(t)
    ldx #t[1]*4 lda #b[1] sta oam,x
    inx lda #b[2] sta oam,x
    inx lda #b[3] sta oam,x
    inx lda #b[4] sta oam,x
end

--[[ button state:
    bit:   	 7     6     5     4     3     2     1     0
    button:	 A     B  Select Start  Up   Down  Left  Right
    https://wiki.nesdev.com/w/index.php/Controller_Reading
]]
-- fast reading of just the A button
-- return C if A is pressed, c otherwise
read_joy_a = function(joy_index)
    local joy = joy_index == 2 and SPECIO2 or SPECIO1
    lda #1 sta joy lsr sta joy lda joy lsr
end
-- read one joypad state into dst
read_joy = function(dst, joy_index)
    local joy = joy_index == 2 and SPECIO2 or SPECIO1
    lda #1 sta joy sta dst lsr sta joy
    @_readbutton lda joy lsr rol dst bcc _readbutton
end
-- read both joypad states and Famicom's DA15 expansion port
read_joys = function(dst1, dst2)
    lda #1 sta SPECIO1 sta dst2 lsr sta SPECIO1
@_readbuttons
    lda SPECIO1 and #3 cmp #1 rol dst1
    lda SPECIO2 and #3 cmp #1 rol dst2
    bcc _readbuttons
end
-- read both joypad states on even cycles only, to safely work with DPCM
-- must be called right after oamcache_flush or any other sta OAMDMA
read_joys_even = function(dst1, dst2)
    ldx #1 stx dst1 stx SPECIO1 dex stx SPECIO1
@_readbuttons
    lda SPECIO2 and #3 cmp #1 rol dst2,x
    lda SPECIO1 and #3 cmp #1 rol dst1
    bcc _readbuttons
end

init = function()
    sei -- cld not needed, no BCD support
    ldx #0x40 stx SPECIO2 -- disable APU frame IRQ
    ldx #0xff txs inx stx PPUCTRL stx PPUMASK stx DMCFREQ -- disable NMI, rendering, DMC IRQs
    bit PPUSTAT -- clear remnant VBlank PPU status flag on reset
    vblank_waitbegin()
    lda #0 sta SNDCNT -- stop APU channels
    -- clear CPU RAM
@_zeroram
    sta 0x0000,x sta 0x0100,x sta 0x0200,x sta 0x0300,x
    sta 0x0400,x sta 0x0500,x sta 0x0600,x sta 0x0700,x
    inx bne _zeroram
    vblank_waitbegin()
    -- clear OAM
    oamcache_clear() oamcache_flush()
    -- clear PPU RAM
    bit PPUSTAT ppu_addr(0x2000) tax ldy #0x10
@_zeroppu
    sta PPUDATA dex bne _zeroppu dey bne _zeroppu
    bit PPUSTAT -- reset latch
    if mappers.init then mappers.init() end
end

-- NES 2.0 (backward compatible with iNES)
-- https://wiki.nesdev.com/w/index.php/NES_2.0
header = function(t)
    if not t then t = {} end
    local logsz = function(sz)
        assert(sz >= 0 and sz <= 1048576, "invalid size: " .. sz .. ", expected [0, 1048576]")
        if sz < 1 then return 0 end
        if sz <= 128 then return 1 end
        return math.ceil(math.log(sz/64, 2))
    end
    -- mapper
    local mi1 = t.mapperid or 0
    assert(mi1 >= 0 and mi1 < 4096, "invalid mapper id: " .. mi1 .. ", expected [0, 4095]")
    local ms1 = t.submapperid or 0
    assert(ms1 >= 0 and ms1 < 16, "invalid submapper id: " .. ms1 .. ", expected [0, 15]")
    local mapper6 = (mi1 & 0xf) << 4
    local mapper7 = mi1 & 0xf0
    local mapper8 = (mi1 >> 8) | (ms1 << 4)
    -- prgsize
    local prgsize = math.tointeger((t.prgsize or 16384) / 16384)
    assert(prgsize, "prgsize must be a multiple of 16384")
    -- chrsize
    local chrsize = math.tointeger((t.chrsize or 0) / 8192)
    assert(chrsize, "chrsize must be a multiple of 8192")
    -- wramsize (not battery-backed)
    local wramsize = logsz(t.wramsize or 0)
    -- bramsize (battery-backed)
    local bramsize = logsz(t.bramsize or 0)
    -- chrbramsize (battery-backed)
    local chrbramsize = logsz(t.chrbramsize or 0)
    -- chrramsize (not battery-backed)
    local chrramsize = logsz(t.chrramsize or (chrbramsize==0 and chrsize==0 and 8192 or 0))
    local battery_bit = bramsize == 0 and chrbramsize == 0 and 0 or 2
    -- mirror: 'H' for horizontal mirroring, 'V' for vertical mirroring
    --         '4' for four-screen VRAM, 218 for four-screen and vertical
    local mirror = (t.mirror or 'h'):lower()
    mirror = ({ h=0, v=1, ['4']=8, [218]=9 })[mirror]
    assert(mirror, "invalid mirror mode: " .. mirror .. ", expected 'H', 'V', '4', or 218")
    -- tv: 'N' for NTSC, 'P' for PAL, 'NP' for both preferring NTSC, 'PN' for both preferring PAL
    local tv, tvm = 0, (t.tv or 'n'):lower()
    assert(tvm=='n' or tvm=='p' or tvm=='np' or tvm=='pn', "invalid tv mode: " .. tostring(t.tv) .. ", expected 'N', 'P', 'NP' or 'PN'")
    if tvm[1] == 'p' then tv = 1 end 
    if #tvm > 1 then tv = tv + 2 end

    @@header -- size: 16 bytes
    dc.b 0x4e, 0x45, 0x53 -- 'NES'
    dc.b 0x1a
    dc.b prgsize, chrsize
    dc.b mapper6 | mirror | battery_bit
    dc.b mapper7 | 8
    dc.b mapper8
    dc.b ((chrsize >> 4) & 0xF0) | ((prgsize >> 8) & 0x0F)
    dc.b (bramsize << 4) | wramsize
    dc.b (chrbramsize << 4) | chrramsize
    dc.b tv, 0, 0, 0

    -- update table with defaulted values
    t.prgsize = prgsize * 16384
    t.chrsize = chrsize * 8192
    t.wramsize = math.tointeger(2^wramsize*64)
    t.bramsize = math.tointeger(2^bramsize*64)
    t.chrbramsize = math.tointeger(2^chrbramsize*64)
    t.chrramsize = math.tointeger(2^chrramsize*64)
    mappers.header=t
end

local n0ne = function(x) return not x or x == 0 end
local val0 = function(x) return x and x or 0 end

mappers.NROM = function(t)
    if not t then t = {} end
    if not t.prgsize then t.prgsize = 16384 end
    assert(t.prgsize == 16384 or t.prgsize == 32768, "prgsize must be 16 or 32kB")
    if n0ne(t.chrsize) and n0ne(t.chrramsize) and n0ne(t.chrbramsize) then t.chrsize = 8192 end
    assert(val0(t.chrsize) + val0(t.chrramsize) + val0(t.chrbramsize) == 8192, "combined chrrom size must be 8kB")
    assert(not t.mirror or ({ h=1, v=1 })[t.mirror:lower()], "only H and V mirroring are supported")
    local prgstart = 0x10000 - t.prgsize
    hdrrom = location{prgstart - 16, prgstart - 1, name='header'}
    header(t)
    prgrom = location{prgstart, 0xffff, name='prgrom'}
    section{"vectors", org=0xfffa} dc.w nmi, main, irq
    if (t.chrsize > 0) then chrrom = location{0x10000, 0x10000 + 0x1fff, name='chrrom'} end
end
mappers[0] = mappers.NROM

mappers.MMC1 = function(t)
end
mappers[1] = mappers.MMC1

--[[
    prgroms are numbered from last (0) to first (#-1), so that adding more does not change
    prgrom0, which must contain the reset vector (main).
    Last 2 prg banks are merged into 1 16kB bank, to allow linker optimization - hence, 0 must
    always be set on bit 6 of bank select, even; also, there is no prgrom1 as a consequence.
    t.prgmap is an optional function taking a prgrom bank index and returning its rorg value.
    Default is to map even banks to 0x8000, and odd to 0xa000.

    chrroms are all 1kB, so they can work with chr A12 inversion enabled or not.

    With default submapper id of 0, this defaults to revision MM3C, which generates a scanline
    interrupt at each scanline when counter is loaded with 0.
]]
mappers.MMC3 = function(t)
    if not t then t = {} end
    t.mapperid = 4
    if not t.bramsize then t.bramsize = 8192 end
    assert(t.bramsize == 8192, "bramsize must be 8kB")
    if not t.prgsize then t.prgsize = 32768 end
    assert(t.prgsize >= 0x8000 and t.prgsize <= 0x80000, "prgsize must be at least 32kB and at most 512kB")
    if not t.chrsize then t.chrsize = 8192 end
    assert(t.chrsize >= 0x2000 and t.chrsize <= 0x40000, "chrsize must be at least 8kB and at most 256kB")
    hdrrom = location{0x7FF0, 0x7FFF, name='header'}
    header(t)
    local prgmap = t.prgmap or function(bi, bc) return 0x8000+(bi&1)*0x2000 end
    local bc = t.prgsize//0x2000
    for bi=0,bc-3 do
        local o,ix = 0x8000 + bi*0x2000, bc-bi-1
        _ENV['prgrom'..ix] = location{o, o+0x1fff, rorg=prgmap(ix,bc), name='prgrom'..ix}
    end
    do
        local o = 0x8000 + (bc-2)*0x2000
        prgrom0 = location{o, o+0x3fff, rorg=0xc000, name='prgrom0'}
        section{"vectors", org=o+0x3ffa} dc.w nmi, main, irq
    end
    local chrstart = 0x8000 + bc*0x2000
    local cc = t.chrsize//0x400
    for ci=0,cc-1 do
        local o = chrstart + ci*0x400
        _ENV['chrrom'..ci] = location{o, o+0x3ff, name='chrrom'..ci}
    end
    function switchprgrom(slot, bankix)
        assert(slot<2)
        lda #6+slot sta 0x8000
        assert(bankix < bc)
        bankix = bc-1-bankix -- reverse index order, since we set 0 as last
        lda #bankix sta 0x8001
    end
    local a12inv = false
    function seta12inv(enabled) a12inv = enabled end
    -- slot [0, 7], each slot is 1kB counting in order, regardless of a12inv state
    function switchchrrom(slot, bankix)
        assert(slot<8)
        if a12inv then 
            assert(slot ~= 5 and slot ~= 7)
            if slot == 6 then slot = 1
            elseif slot == 4 then slot = 0
            else slot = slot + 2 end
            slot = slot | 0x80
        else
            assert(slot ~= 1 and slot ~= 3)
            if slot == 2 then slot = 1
            elseif slot > 3 then slot = slot - 2 end
        end
        lda #slot sta 0x8000
        assert(bankix < cc)
        lda #bankix sta 0x8001
    end
    function setmirror(mirror)
        mirror = assert(({ h=1, v=0 })[mirror:lower()])
        lda #mirror sta 0xa000
    end
    function protectsram() lda 0x40 sta 0xa001 end
    function scanlineirq(count) ldx #1 stx 0xe000 lda #count-1 sta 0xc000 sta 0xc001 stx 0xe001 end
    mappers.init = function()
        switchprgrom(0, 2) switchprgrom(1, 3)
        switchchrrom(0, 0) switchchrrom(2, 2)
        switchchrrom(4, 4) switchchrrom(5, 5) switchchrrom(6, 6) switchchrrom(7, 7)
        local mirror = mappers.header.mirror
        if mirror==0 or mirror==1 then lda #mirror~1 sta 0xa000 end
        lda #0x80 sta 0xa001
    end
end
mappers[4] = mappers.MMC3