$bytes = []

def zero_fill(pos)
	while ($bytes.length <= pos) do
		$bytes.push(0)
	end
end

def word(i)
	$bytes.push((i & 0x000000ff),
		(i & 0x0000ff00) >> 8,
		(i & 0x00ff0000) >> 16,
		(i & 0xff000000) >> 24)
end

def string(s)
	$bytes.push *s.bytes
	while (($bytes.length % 4) != 0) do
		$bytes.push 0
	end
end


# # # REGISTER NAMES # # #
def r0() 0 end # zero
def at() 1 end # assembler temporary

# subroutine return values, may be changed by subroutines
def v0() 2 end ; def v0() 3 end

# subroutine arguments, may be changed by subroutines
def a0() 4 end ; def a1() 5 end
def a2() 6 end ; def a3() 7 end

# temporaries, may be changed by subroutines
def t0() 8 end ; def t1() 9 end
def t2() 10 end ; def t3() 11 end
def t4() 12 end ; def t5() 13 end
def t6() 14 end ; def t7() 15 end

# static variables, must be saved by subs
def s0() 16 end ; def s1() 17 end
def s2() 18 end ; def s3() 19 end
def s4() 20 end ; def s5() 21 end
def s6() 22 end ; def s7() 23 end

# temporaries, may be changed by subroutines
def t8() 24 end ; def t9() 25 end

# reserved for kernel, detroyed by some irq handlers
def k0() 26 end ; def k1() 27 end

def gp() 28 end # global pointer, rarely used
def sp() 29 end # stack pointer
def fp() 30 end # frame pointer
def ra() 31 end # return address



# # # INSTRUCTIONS # # #

# ops should always be in format of
# [src] [args] -> [dest]

def jmp(addr)
	# TODO: this probs wont show up since we have < 3mb of ram but
	# we can't jump to an address with a different 4 most significant bits
	raise "jmp addr must be 4 byte aligned" if addr % 4 != 0
	op = 0b0000_1000_0000_0000_0000_0000_0000_0000
	op = op | ((addr >> 2) & 0b11_11111111_11111111_11111111)
	word op
end

def jal(addr)
	# TODO: same problem as jmp
	raise "jmp addr must be 4 byte aligned" if addr % 4 != 0
	op = 0b0000_1100_0000_0000_0000_0000_0000_0000
	op = op | ((addr >> 2) & 0b11_11111111_11111111_11111111)
	word op
end

# load upper immediate
# NOTE: val will be shifted left by 16
def lui(val, dest)
	#      oooooo ----- ddddd iiiiiiiiiiiiiiii
	op = 0b001111_00000_00000_0000000000000000
	op |= 0xffff & val
	op |= (0b11111 & dest) << 16
	word op
end

# shift right logical
def srl(src, amt, dest)
	#      oooooo ----- sssss ddddd hhhhh oooooo
	op = 0b000000_00000_00000_00000_00000_000010
	op |= (src & 0b11111) << 16
	op |= (amt & 0b11111) << 6
	op |= (dest & 0b11111) << 11
	word op
end

# or immediate
def ori(src, bits, dest)
	#      oooooo sssss ddddd iiiiiiii iiiiiiii
	op = 0b001101_00000_00000_00000000_00000000
	op |= (src & 0b11111) << 21
	op |= (bits & 0xffff)
	op |= (dest & 0b11111) << 16
	word op
end

# store word
# NOTE: this doesn't work unaligned
def sw(src, offset, dest)
	#      oooooo ddddd sssss iiiiiiii iiiiiiii
	op = 0b101011_00000_00000_00000000_00000000
	op |= (src & 0b11111) << 16
	op |= (offset & 0xffff)
	op |= (dest & 0b11111) << 21
	word op
end

# add immediate unsigned
def addiu(src, val, dest)
	#      oooooo sssss ddddd iiiiiiii iiiiiiii
	op = 0b001001_00000_00000_00000000_00000000
	op |= (src & 0b11111) << 21
	op |= (val & 0xffff)
	op |= (dest & 0b11111) << 16
	word op
end

# add immediate
def addi(src, val, dest)
	#      oooooo sssss ddddd iiiiiiii iiiiiiii
	op = 0b001000_00000_00000_00000000_00000000
	op |= (src & 0b11111) << 21
	op |= (val & 0xffff)
	op |= (dest & 0b11111) << 16
	word op
end


# # # PSEUDO-INSTRUCTIONS # # #
def nop
	word 0x00000000
end

# load word immediate
def lwi(val, dest)
	lui val >> 16, dest
	ori dest, val, dest
end





def label
	return $bytes.length + $base_addr
end

# this is a terrible idea and im doing it
#$labels = {}
#def method_missing(m, *args)
#	puts m.to_s
#	super
#end


# # # CONSTANTS # # #
$base_addr = 0x80010000 # 0x10000 # 0x80010000
$file_size = 0x800 * 3
$gp0 = 0x1F801810
$gp1 = 0x1F801814

# # # EXE HEADER # # #
string "PS-X EXE"
zero_fill 0x00f
word $base_addr # initial pc
word 0x00000000 # initial GP/R28
word $base_addr # destintation address in RAM
word $file_size # file size excluding header (must be N * 0x800)
word 0x00000000 # Unknown/Unused
word 0x00000000 # Unknown/Unused
word 0x00000000 # Memfill start address
word 0x00000000 # Memfill size in bytes
word 0x801ffff0 # Initial SP/R29 & FP/R30 Base
word 0x00000000 # Initial SP/R29 & FP/R30 Offs
zero_fill 0x4b # Reserved for A(43h) Function
# Ascii marker would go here
zero_fill 0x7ff
$exe_header = $bytes[0..]
$bytes = []




# # # PROGRAM CODE # # #
nop

# ops should always be in format of
# src [args] -> dest

# THIS IS NEEDED
lwi $gp1, t0
lwi 0x03_00_00_00, t1 # display enable
sw t1, 0, t0

lwi $gp0, t0

lwi 0xe1_000000 + 0b0_0_0_1_0_01_00_0_0000, t1
sw t1, 0, t0

# THIS IS NEEDED
lwi 0xe4_000000 + (640) + (480 << 10), t1 # Drawing Area bottom right
sw t1, 0, t0

def line(st, en)
	lwi 0x40_ffffff, t1 # monochrome line, color ffffff
	sw t1, 0, t0

	lwi st, t1 # line vert 1
	sw t1, 0, t0

	lwi en, t1 # line vert 2
	sw t1, 0, t0
end

# l_end_loop = label
line 0x0010_0010, 0x0030_0030
line 0x0030_0030, 0x0060_0000
line 0x0060_000a, 0x00f0_0060

nop ; nop ; nop ; nop
nop ; nop ; nop ; nop
nop ; nop ; nop ; nop
nop ; nop ; nop ; nop

l_end_loop = label

addiu r0, -10, t3
#addiu t3, 6, t3
addiu t3, -10, t3
nop

addi r0, -10, t4
#addi t4, 6, t4
addi t4, -10, t4
nop

nop
jmp l_end_loop
nop






puts "assembled 0x#{$bytes.length.to_s 16} (#{$bytes.length}) bytes"
raise "CODE IS GREATER THAN FILESIZE" if $bytes.length > $file_size
zero_fill $file_size - 1

f = File.new "LOADTHIS.EXE", "wb"
f.write $exe_header.pack("C*")
f.write $bytes.pack("C*")
f.close