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

vcs = {
    -- TIA write only
    VSYNC   = 0x00, -- ......1.  vertical sync set-clear
    VBLANK  = 0x01, -- 11....1.  vertical blank set-clear
    WSYNC   = 0x02, -- <strobe>  wait for leading edge of horizontal blank
    RSYNC   = 0x03, -- <strobe>  reset horizontal sync counter
    NUSIZ0  = 0x04, -- ..111111  number-size player-missile 0
    NUSIZ1  = 0x05, -- ..111111  number-size player-missile 1
    COLUP0  = 0x06, -- 1111111.  color-lum player 0 and missile 0
    COLUP1  = 0x07, -- 1111111.  color-lum player 1 and missile 1
    COLUPF  = 0x08, -- 1111111.  color-lum playfield and ball
    COLUBK  = 0x09, -- 1111111.  color-lum background
    CTRLPF  = 0x0a, -- ..11.111  control playfield ball size & collisions
    REFP0   = 0x0b, -- ....1...  reflect player 0
    REFP1   = 0x0c, -- ....1...  reflect player 1
    PF0     = 0x0d, -- 1111....  playfield register byte 0
    PF1     = 0x0e, -- 11111111  playfield register byte 1
    PF2     = 0x0f, -- 11111111  playfield register byte 2
    RESP0   = 0x10, -- <strobe>  reset player 0
    RESP1   = 0x11, -- <strobe>  reset player 1
    RESM0   = 0x12, -- <strobe>  reset missile 0
    RESM1   = 0x13, -- <strobe>  reset missile 1
    RESBL   = 0x14, -- <strobe>  reset ball
    AUDC0   = 0x15, -- ....1111  audio control 0
    AUDC1   = 0x16, -- ....1111  audio control 1
    AUDF0   = 0x17, -- ...11111  audio frequency 0
    AUDF1   = 0x18, -- ...11111  audio frequency 1
    AUDV0   = 0x19, -- ....1111  audio volume 0
    AUDV1   = 0x1a, -- ....1111  audio volume 1
    GRP0    = 0x1b, -- 11111111  graphics player 0
    GRP1    = 0x1c, -- 11111111  graphics player 1
    ENAM0   = 0x1d, -- ......1.  graphics (enable) missile 0
    ENAM1   = 0x1e, -- ......1.  graphics (enable) missile 1
    ENABL   = 0x1f, -- ......1.  graphics (enable) ball
    HMP0    = 0x20, -- 1111....  horizontal motion player 0
    HMP1    = 0x21, -- 1111....  horizontal motion player 1
    HMM0    = 0x22, -- 1111....  horizontal motion missile 0
    HMM1    = 0x23, -- 1111....  horizontal motion missile 1
    HMBL    = 0x24, -- 1111....  horizontal motion ball
    VDELP0  = 0x25, -- .......1  vertical delay player 0
    VDELP1  = 0x26, -- .......1  vertical delay player 1
    VDELBL  = 0x27, -- .......1  vertical delay ball
    RESMP0  = 0x28, -- ......1.  reset missile 0 to player 0
    RESMP1  = 0x29, -- ......1.  reset missile 1 to player 1
    HMOVE   = 0x2a, -- <strobe>  apply horizontal motion
    HMCLR   = 0x2b, -- <strobe>  clear horizontal motion registers
    CXCLR   = 0x2c, -- <strobe>  clear collision latches= 0x20

    -- TIA read only
    CXM0P   = 0x30, -- 11......  read collision M0-P1, M0-P0 (Bit 7 --6)
    CXM1P   = 0x31, -- 11......  read collision M1-P0, M1-P1
    CXP0FB  = 0x32, -- 11......  read collision P0-PF, P0-BL
    CXP1FB  = 0x33, -- 11......  read collision P1-PF, P1-BL
    CXM0FB  = 0x34, -- 11......  read collision M0-PF, M0-BL
    CXM1FB  = 0x35, -- 11......  read collision M1-PF, M1-BL
    CXBLPF  = 0x36, -- 1.......  read collision BL-PF, unused
    CXPPMM  = 0x37, -- 11......  read collision P0-P1, M0-M1
    INPT0   = 0x38, -- 1.......  read pot port
    INPT1   = 0x39, -- 1.......  read pot port
    INPT2   = 0x3a, -- 1.......  read pot port
    INPT3   = 0x3b, -- 1.......  read pot port
    INPT4   = 0x3c, -- 1.......  read input
    INPT5   = 0x3d, -- 1.......  read input

    -- PIA
    SWCHA   = 0x280, -- 11111111  Port A= 0x280 -- input or output  (read or write)
    SWACNT  = 0x281, -- 11111111  Port A DDR, 0= input, 1=output
    SWCHB   = 0x282, -- 11111111  Port B= 0x280 -- console switches (read only)
    SWBCNT  = 0x283, -- 11111111  Port B DDR (hardwired as input)
    INTIM   = 0x284, -- 11111111  Timer output (read only)
    INSTAT  = 0x285, -- 11......  Timer Status (read only, undocumented)

    TIM1T   = 0x294, -- 11111111  set 1 clock interval (838 nsec/interval)
    TIM8T   = 0x295, -- 11111111  set 8 clock interval (6.7 usec/interval)
    TIM64T  = 0x296, -- 11111111  set 64 clock interval (53.6 usec/interval)
    T1024T  = 0x297, -- 11111111  set 1024 clock interval (858.2 usec/interval)
}
do
    local symbols = cpu.symbols
    for k,v in pairs(vcs) do symbols[k] = v end
end

nusiz = {
    MSL_SIZE_1 = 16*0,
    MSL_SIZE_2 = 16*1,
    MSL_SIZE_4 = 16*2,
    MSL_SIZE_8 = 16*3,
    ONE_COPY = 0,
    TWO_COPIES_CLOSE = 1,
    TWO_COPIES_MEDIUM = 2,
    THREE_COPIES_CLOSE = 3,
    TWO_COPIES_WIDE = 4,
    DOUBLE_SIZED_PLAYER = 5,
    THREE_COPIES_MEDIUM = 6,
    QUAD_SIZED_PLAYER = 7,
}

ctrlpf = {
	BALL_SIZE_1 = 16*0,
	BALL_SIZE_2 = 16*1,
	BALL_SIZE_4 = 16*2,
	BALL_SIZE_8 = 16*3,
	PF_MIRRORED = 1,
	PF_SCOREMODE = 2,
	PF_PRIO_BALLABOVE = 4,
}

enable = {
	DISABLE = 0,
	ENABLE = 2,
}

vdel = {
	DISABLE = 0,
	ENABLE = 1,
}

hm = {
	LEFT_7 = 7*16,
	LEFT_6 = 6*16,
	LEFT_5 = 5*16,
	LEFT_4 = 4*16,
	LEFT_3 = 3*16,
	LEFT_2 = 2*16,
	LEFT_1 = 1*16,
	NO_MOTION = 0,
	RIGHT_1 = 15*16,
	RIGHT_2 = 14*16,
	RIGHT_3 = 13*16,
	RIGHT_4 = 12*16,
	RIGHT_5 = 11*16,
	RIGHT_6 = 10*16,
	RIGHT_7 = 9*16,
	RIGHT_8 = 8*16,
	
	C74_LEFT_15 = 7*16,
	C74_LEFT_14 = 6*16,
	C74_LEFT_13 = 5*16,
	C74_LEFT_12 = 4*16,
	C74_LEFT_11 = 3*16,
	C74_LEFT_10 = 2*16,
	C74_LEFT_9 = 1*16,
	C74_LEFT_8 = 0*16,
	C74_LEFT_7 = 15*16,
	C74_LEFT_6 = 14*16,
	C74_LEFT_5 = 13*16,
	C74_LEFT_4 = 12*16,
	C74_LEFT_3 = 11*16,
	C74_LEFT_2 = 10*16,
	C74_LEFT_1 = 9*16,
	C74_NO_MOTION = 8*16,
}

NTSC = {
    TIM_OVERSCAN    = 38, -- TIM64T,  2432 cycles =   32 scanlines
    TIM_VBLANK      = 45, -- TIM64T,  2880 cycles = ~ 38 scanlines
    TIM_KERNEL      = 15, -- T1024T, 15360 cycles = ~202 scanlines
}
PAL = {
    TIM_OVERSCAN    = 50, -- TIM64T,  3200 cycles = ~ 42 scanlines
    TIM_VBLANK      = 61, -- TIM64T,  3904 cycles = ~ 51 scanlines
    TIM_KERNEL      = 17, -- T1024T, 17408 cycles = ~229 scanlines
}
TV = PAL

init = function() cld ldx#0 txa local l=label() dex tsx pha bne l end
wait = function() local l=label() lda INTIM bne l end
overscan_begin = function() sta WSYNC lda#2 sta VBLANK lda#TV.TIM_OVERSCAN sta TIM64T end
overscan_end = wait
overscan = function(f) overscan_begin() if f then f() end overscan_end() end
vblank_begin = function()
    lda#0b1110 local l=label() sta WSYNC sta VSYNC lsr bne l
    lda#2 sta VBLANK lda#TV.TIM_VBLANK sta TIM64T
end
vblank_end = function() wait() sta WSYNC sta VBLANK end
vblank = function(f) vblank_begin() if f then f() end vblank_end() end
screen_begin = function() lda#TV.TIM_KERNEL sta T1024T end
screen_end = wait
screen = function(f) screen_begin() if f then f() end screen_end() end

setptr = function(f, add) if not add then add=tmp end lda#op_resolve(f)&0xff sta add lda#op_resolve(f)>>8 sta add+1 end
xcall = function(f) setptr(f, mappers.tmp) jsr jmpfar end
rtximp = function() jmp rtsfar end

-- for mappers that swap banks in place
-- call an asm function into another bank and generate the stubs if needed
local far_stubs = {} _ENV.far_stubs=far_stubs
farcall = function(dst)
    local loc_src = location_current
    local create_stubs = function()
        if type(dst) == 'function' then dst = dst() end
        local stub,loc_dst = far_stubs[dst],dst.section.location
        if not (stub and stub[loc_src]) then
            location(loc_src)
            local stub_caller = section()   switchrom(loc_dst.rom) skip(6) rts
            location(loc_dst)
            local stub_callee = section()   jsr dst switchrom(loc_src.rom)
            relate(stub_caller, stub_callee, 3)
            if not stub then far_stubs[dst] = {} end
            far_stubs[dst][loc_src] = stub_caller
        end
    end
    -- create stubs at the end to get to the referenced labels
    table.insert(before_link, create_stubs)
    jsr far_stubs[dst][loc_src]
end

mappers = { tmp=0x8a }

mappers['2K'] = function()
    rom0 = location(0xf800, 0xffff)
    section{"vectors", org=0xfffc} dc.w main if mappers.irq then dc.w main end
end
mappers.CV = mappers['2K']

mappers['4K'] = function()
    rom0 = location(0xf000, 0xffff)
    section{"vectors", org=0xfffc} dc.w main if mappers.irq then dc.w main end
end

local bank_stubs = function(count, hotspot, entry)
    function switchrom(i) assert(i>=0 and i<count) if mappers.noillegal then bit 0x1000+hotspot+i else nop 0x1000+hotspot+i end end
    local base = 0x10000 - (count << 12)
    local jmpfar0,rtsfar0
    for bi=0,count-1 do
        local o,ro = base+(bi<<12), mappers.jmpfar and 0x1000+(bi<<13) or 0xf000 -- stella doesn't honor breakpoints in banks with jmpfar method
        _ENV['rom' .. bi] = location{o, o+0xfff, rorg=ro, rom=bi}
        local start=section{"entry"..bi, org=o+(entry or hotspot-6)} switchrom(0) if bi==0 then jmp main end
        section{"switchtab"..bi, org=o+hotspot} for i=1,count do byte(0) end
        section{"vectors"..bi, org=o+0xffc} dc.w start if mappers.irq then dc.w start end
        if mappers.jmpfar then
            local jmpfar = section("jmpfar"..(bi>0 and bi or ''))
                lda mappers.tmp+1 lsr lsr lsr lsr lsr tax
                if mappers.noillegal then and 0x1000+hotspot,x else nop 0x1000+hotspot,x end
                jmp (mappers.tmp)
            local rtsfar = section("rtsfar"..(bi>0 and bi or ''))
                tsx lda 2,x lsr lsr lsr lsr lsr tax
                if mappers.noillegal then and 0x1000+hotspot,x else nop 0x1000+hotspot,x end
                rts
            if bi==0 then jmpfar0=jmpfar rtsfar0=rtsfar
            else relate(jmpfar0, jmpfar) relate(rtsfar0, rtsfar) end
        end
    end
end
mappers.FA = function() bank_stubs(3, 0xff8) end
mappers.F8 = function() bank_stubs(2, 0xff8) end
mappers.F6 = function() bank_stubs(4, 0xff6) end
mappers.F4 = function() bank_stubs(8, 0xff4) end
mappers.EF = function() bank_stubs(16, 0xfe0, 0xffc-6) end

mappers.FE = function()
    rom1 = location(0xd000, 0xdfff)
    section{"vectors1", org=0xdffc} dc.w main if mappers.irq then dc.w main end
    location{0xe000, 0xefff, nofill=true}
    rom0 = location(0xf000, 0xffff)
    section{"vectors0", org=0xfffc} dc.w main if mappers.irq then dc.w main end
end

mappers.E0 = function(map)
    function switchrom0(i) assert(map[i]==0) if mappers.noillegal then bit 0x1fe0+i else nop 0x1fe0+i end end
    function switchrom1(i) assert(map[i]==1) if mappers.noillegal then bit 0x1fe8+i else nop 0x1fe8+i end end
    function switchrom2(i) assert(map[i]==2) if mappers.noillegal then bit 0x1ff0+i else nop 0x1ff0+i end end
    map[7]=3
    for bi=0,7 do
        local o = bi*0x400
        _ENV["rom"..bi] = location{0xe000+o, 0xe3ff+o, rorg=0xf000+map[bi]*0x400}
    end
    section{"switchtab", org=0xffe0} for i=1,8*3 do byte(0) end
    section{"vectors", org=0xfffc} dc.w main if mappers.irq then dc.w main end
end

-- rom0 refers to the last one in ROM, ie the one always mapped to 0xF800-0xFFFF,
-- such that changing the rom count does not change its index
mappers['3F'] = function(count)
    function switchrom(i) assert(i>=0 and i<count-1) lda#i sta 0x3f end
    local symbols=cpu.symbols for k,v in pairs(vcs) do -- remap TIA to 0x40-0x7F
        if v<0x40 then v=v+0x40 vcs[k]=v symbols[k]=v end
    end
    bi=0,count-2 do
        local o = 0x800*bi + 0x10000
        _ENV["rom"..(count-1-bi)] = location{o, 0x7ff+o, rorg=0xf000}
    end
    local o = 0x800 * (count-1) + 0x10000
    rom0 = location{o, 0x7ff+o, rorg=0xf800}
    section{"vectors", org=0xfffc} dc.w main if mappers.irq then dc.w main end
end

mappers.E7 = function()
    function switchrom(i) assert(i>=0 and i<7) if mappers.noillegal then bit 0x1fe0+i else nop 0x1fe0+i end end
    function enableram() if mappers.noillegal then bit 0x1fe7 else nop 0x1fe7 end end
    function switchram(i) assert(i>=0 and i<4) if mappers.noillegal then bit 0x1fe8+i else nop 0x1fe8+i end end
    for bi=0,6 do
        local o = bi*0x800
        _ENV['rom' .. bi] = location{o+0xc000, o+0xc7ff, rorg=0xf000}
    end
    rom7 = location{0xf800, 0xffff, rorg=0xf800}
    section{"switchtab", org=0xffe0} for i=1,12 do byte(0) end
    section{"vectors", org=0xfffc} dc.w main if mappers.irq then dc.w main end
end

mappers.F0 = function()
    function switchrom() lda 0x1ff0 end
    for bi=0,15 do
        local o = bi*0x1000 + 0x1000
        _ENV['rom' .. bi] = location{o, o+0xfff, rorg=0xf000}
        local start=section{"entry"..bi, org=o+0xffc-8}
            local l=label() switchrom() bne l
            if bi==0 then jmp main end
        section{"switchtab"..bi, org=o+0xff0} byte(bi)
        section{"vectors"..bi, org=o+0xffc} dc.w start if mappers.irq then dc.w start end
    end
end

local bank_stubs2 = function(hotspot0, hotspot1)
    function switchrom(i)
        if i==0 then if mappers.noillegal then bit hotspot0 else nop hotspot0 end
        elseif i==1 then if mappers.noillegal then bit hotspot1 else nop hotspot1 end
        else error("invalid rom index: " .. i) end
    end
    local base = 0xe000
    for bi=0,1 do
        local o = base+(bi<<12)
        _ENV['rom' .. bi] = location{o, o+0xfff, rorg=0xf000}
        local start=section{"entry"..bi, org=o+0xffc-6} switchrom(0) if bi==0 then jmp main end
        section{"vectors"..bi, org=o+0xffc} dc.w start if mappers.irq then dc.w start end
    end
end
mappers.UA = function() bank_stubs2(0x220, 0x240) end
mappers['0840'] = function() bank_stubs2(0x800, 0x840) end
mappers.SB = function(count)
    function switchrom(i) assert(i>=0 and i<count) if mappers.noillegal then bit 0x800+i else nop 0x800+i end end
    for bi=0,count-1 do
        local o = bi*0x1000 + 0x1000
        _ENV['rom' .. bi] = location{o, o+0xfff, rorg=0xf000}
        local start=section{"entry"..bi, org=o+0xffc-6} switchrom(0) if bi==0 then jmp main end
        section{"vectors"..bi, org=o+0xffc} dc.w start if mappers.irq then dc.w start end
    end
end

mappers['3E'] = function(rom_count, ram_count)
    mappers['3F'](rom_count)
    function switchram(i) assert(i>=0 and i<ram_count-1) lda#i sta 0x3E end
end

-- rom0 refers to the last 1k in ROM, which is mapped automatically on reset,
-- such that changing the rom count does not change its index
mappers.MC = function(rom_count, ram_count)
    for i=0,3 do
        _ENV['switchrom'..i] = function(i) assert(i>=0 and i<128) lda#i sta 0x3c+i end
        _ENV['switchram'..i] = function(i) assert(i>=0 and i<128) lda#i+0x80 sta 0x3c+i end
    end
    for bi=0,rom_count-1 do
        local o = bi*0x400 + 0x1000
        _ENV["rom"..(count-1-bi)] = location{o, 0x3ff+o, rorg=0xf000+(o&0xfff)}
    end
    local start=section("entry") switchrom3(0) jmp main
    section{"vectors", org=0xfffc} dc.w start if mappers.irq then dc.w start end
end

mappers.X07 = function()
    function switchrom(i) assert(i>=0 and i<16) if mappers.noillegal then bit 0x80b+(i<<4) else nop 0x80b+(i<<4) end end
    -- map TIA also to 0x40-0x7F: use VSYNC for bank 14, and VSYNC2 for bank 15
    vcs.VSYNC2=0x40 cpu.symbols.VSYNC2=0x40
    for bi=0,count-1 do
        local o = bi*0x1000 + 0x1000
        _ENV['rom' .. bi] = location{o, o+0xfff, rorg=0xf000}
        local start=section{"entry"..bi, org=o+0xffc-6} switchrom(0) if bi==0 then jmp main end
        section{"vectors"..bi, org=o+0xffc} dc.w start if mappers.irq then dc.w start end
    end
end

--------------------------------------------------------------------------------
-- Image converters
--------------------------------------------------------------------------------

image_scan = function(img, opt)
    local r, b, flipx, flipy, xinc, yinc, x0, y0, x1, y1, width, height = {}, {}
    xinc=opt.xinc or 1 x0=opt.x0 or 0 x1=opt.x1 or img.width-1
    yinc=opt.yinc or 1 y0=opt.y0 or 0 y1=opt.y1 or img.height-1
    if x1<x0 and xinc<0 then x0,x1=x1,x0 xinc=-xinc flipx=true end
    if y1<y0 and yinc<0 then y0,y1=y1,y0 yinc=-yinc flipy=true end
    assert(y1>=y0 and x1>=x0 and x1<img.width and y1<img.height,
        string.format("invalid dimensions: x[%d -> %d] y[%d -> %d] for img[%d, %d]", x0, x1, y0, y1, img.width, img.height))
    width  = 1 + (x1 - x0) // xinc
    height = 1 + (y1 - y0) // yinc
    for y=y0,y1,yinc do for x=x0,x1,xinc do b[#b+1] = img[y*img.width+x+1] end end
    if flipx then for y=0,height-1 do for x=0,width//2-1 do
        local i0, i1 = y*width+x+1, y*height+width
        b[i0], b[i1] = b[i1], b[i0]
    end end end
    if flipy then for y=0,height//2-1 do for x=0,width-1 do
        local i0, i1 = y*width+x+1, (height-1-y)*width+x+1
        b[i0], b[i1] = b[i1], b[i0]
    end end end
    return b, width, height
end

linecol = function(opt)
    if type(opt) ~= 'table' then opt = { opt } end
    local image = opt.image or assert(l65.image(opt.filename or opt[1]))
    local b,w,h = image_scan(image, opt)
    local lc = {}
    for y=0,h-1 do
        lc[#lc+1] = 0
        for x=1,w do
            local col = b[y*w + x]
            if col ~= 0 then lc[#lc]=col break end
        end
    end
    return lc
end

playfield = function(opt)
    if type(opt) ~= 'table' then opt = { opt } end
    local image = opt.image or assert(l65.image(opt.filename or opt[1]))
    local b,w,h = image_scan(image, opt)
    local pf={} for i=1,6 do pf[i]={} end
    for y=1,h do
        local x = \o,s((b[(y-1)*w+o+1]==0 and 0 or 1)<<s)
        pf[1][y] = x( 0,4)|x( 1,5)|x( 2,6)|x( 3,7)
        pf[2][y] = x( 4,7)|x( 5,6)|x( 6,5)|x( 7,4)|x( 8,3)|x( 9,2)|x(10,1)|x(11,0)
        pf[3][y] = x(12,0)|x(13,1)|x(14,2)|x(15,3)|x(16,4)|x(17,5)|x(18,6)|x(19,7)
        pf[4][y] = x(20,4)|x(21,5)|x(22,6)|x(23,7)
        pf[5][y] = x(24,7)|x(25,6)|x(26,5)|x(27,4)|x(28,3)|x(29,2)|x(30,1)|x(31,0)
        pf[6][y] = x(32,0)|x(33,1)|x(34,2)|x(35,3)|x(36,4)|x(37,5)|x(38,6)|x(39,7)
    end
    if opt.reverse then
        for _,col in ipairs(pf) do for y=1,h do col[y] = col[y] ~ 0xff end end
    end
    return pf
end

sprite = function(opt)
    if type(opt) ~= 'table' then opt = { opt } end
    local image = opt.image or assert(l65.image(opt.filename or opt[1]))
    local b,w,h = image_scan(image, opt)
    local sp={}
    for c=0,((w+7)//8-1) do
        local s={} sp[#sp+1]=s
        for y=0,h-1 do
            local o=c*8
            local e=o+7 >= w and w-1 or o+7
            local i,color = 7,0
            for x=o,e do
                if b[y*w+x+1] ~= 0 then color = color | 1<<i end
                i = i-1
            end
            s[#s+1] = color
        end
    end
    return sp
end