# **************************************************************
#   crc's _ _
#        (_) | ___
#        | | |/ _ \  a tiny virtual computer
#        | | | (_) | 64kw RAM, 32-bit, Dual Stack, MISC
#        |_|_|\___/  ilo.py (c) charles childers
#                           (c) arland childers
# **************************************************************

#Ilo.py v2.2b

#Changes:
#  Split debug modes:
#    'instrs':  Counts overall time, and counts each instruction/bundle.
#    'timing': Just gives elapsed time.
#    'cProfile': Using cProfile, measure time & time for all instructions+overhead. Does not count time for unoptimised bundles.
#  Added more optimised instruction bundles, focusing on DTC.
#  Optimised many instructions, removing uneeded vars. Added raw_lit to not limit it when it is not needed.
#  Faster block unpacking. The uncached blocks seem to only be ~1.7x slower instead of ~2x, now.
#  Sorted some of the instructions listing, in the hopes it may make an improvement. It did not seem to.
#  Added calls to limit the rom and blocks as they get put into memory.


#--- CONFIG
BLOCK_FILE = 'ilo.blocks'
ROM_FILE = 'ilo.rom'
BLOCK_CACHE  = True

DEBUG = False
#  False for nothing
#  'timing' just gives elapsed time. Negligible time loss
#  'Instrs' counts instructions & bundles, but ~2.3x slower
#  'cProfile' for cProfile, but ~3.25x slower


#--- ILO

import os

def limit(val):
  val &= 0xffffffff
  if val & 0x80000000:
    val -= 0x100000000
  return val

def div_mod(a, b):
  x = abs(a)
  y = abs(b)
  q, r = divmod(x, y)
  if a < 0 and b < 0:
    r *= -1
  elif a > 0 and b < 0:
    q *= -1
  elif a < 0 and b > 0:
    r *= -1
    q *= -1
  return q, r


class Stack:
  def __init__(self, size):
    self.sp = 0
    self.data = [0] * size
    self.max = size

  def lit(self, val):
    self.sp += 1
    self.data[self.sp] = ((val & 0xffffffff)-0x100000000) if val & 0x80000000 else (val & 0xffffffff)
  
  def raw_lit(self, val):
    self.sp += 1
    self.data[self.sp] = val
  
  def pop(self):
    self.sp -= 1
    return self.data[self.sp + 1]

  def top(self):
    return self.data[self.sp]

  def second(self):
    return self.data[self.sp - 1]

  def depth(self):
    return self.sp


class BlockTools_NoCache:
  def __init__(self, block_file, rom_file, data_stack, memory):
    self.BLOCK_FILE = block_file
    self.ROM_FILE = rom_file
    self.data_stack = data_stack
    self.memory = memory

  def block_unpack(self, block):
    return [(limit(block[i] | (block[i + 1] << 8) | (block[i + 2] << 16) | (block[i + 3] << 24))) for i in range(0, len(block), 4)]

  def block_pack(self, *values):
    return b''.join(val.to_bytes(4, 'little', signed=True) for val in values)

  def read_block(self):
    buffer = self.data_stack.pop()
    block = self.data_stack.pop()
    with open(self.BLOCK_FILE, 'r+b') as f:
      f.seek(4096 * block)
      self.memory[buffer: buffer + 1024] = list(self.block_unpack(f.read(4096)))

  def block_write(self):
    buffer = self.data_stack.pop()
    block = self.data_stack.pop()
    with open(self.BLOCK_FILE, 'r+b') as f:
      f.seek(4096 * block)
      f.write(self.block_pack(*self.memory[buffer: buffer + 1024]))

  def load_image(self):
    with open(self.ROM_FILE, 'rb') as f:
      raw_data = f.read(4 * 65536)
      self.memory[:] = [limit(int.from_bytes(raw_data[i:i + 4], 'little')) for i in range(0, len(raw_data), 4)]

  def save_image(self):
    with open(self.ROM_FILE, 'wb') as f:
      for value in self.memory[:65536]:
        f.write(value.to_bytes(4, 'little'))
  


class BlockTools(BlockTools_NoCache): # caches blocks.
  def __init__(self, block_file, rom_file, data_stack, memory):
    self.BLOCK_FILE = block_file
    self.ROM_FILE = rom_file
    self.data_stack = data_stack
    self.memory = memory
    self.block_time = os.path.getmtime(block_file)
    self.blocks = []
  
  def read_block(self):
    if self.block_time != os.path.getmtime(self.BLOCK_FILE):
      self.read_all_blocks()
      self.block_time = os.path.getmtime(self.BLOCK_FILE)
    buffer = self.data_stack.pop()
    self.memory[buffer: buffer + 1024] = self.blocks[self.data_stack.pop()]

  def read_all_blocks(self):
    with open(self.BLOCK_FILE, 'rb') as f:
      all_block_data = f.read()
    self.blocks = [list(self.block_unpack(all_block_data[i:i+4096])) for i in range(0, len(all_block_data), 4096)]
    return self.blocks


class ilo:
  def __init__(self, block_file='ilo.blocks', rom_file='ilo.rom', cache_blocks=False):
    self.ip = 0
    self.memory = [0]
    self.input_buffer = []
    self.data = Stack(33)
    self.address = Stack(257)
    if cache_blocks:
      self.block_tools = BlockTools(block_file, rom_file, self.data, self.memory)
    else:
      self.block_tools = BlockTools_NoCache(block_file, rom_file, self.data, self.memory)
    self.instructions = {
      1: self.i01,
      4: self.i04,
      11: self.i11,
      6: self.i06,
      2: self.i02,
      5: self.i05,
      18: self.i18,
      2049: self.i2049,
      2832: self.i2832,
      10: self.i10,
      68288770: self.i68288770,
      8: self.i08,
      17563906: self.i17563906,
      7: self.i07,
      524545: self.i524545,
      3: self.i03,
      459014: self.i459014,
      9: self.i09,
      12: self.i12,
      13: self.i13,
      14: self.i14,
      15: self.i15,
      16: self.i16,
      17: self.i17,
      19: self.i19,
      20: self.i20,
      21: self.i21,
      22: self.i22,
      23: self.i23,
      24: self.i24,
      25: self.i25,
      26: self.i26,
      27: self.i27,
      28: self.i28,
      29: self.i29,
      1793: self.i1793,
      459009: self.i459009,
      67502597: self.i67502597,
      100926722: self.i100926722,
      302059522: self.i302059522,
      2818: self.i2818,
      219156993: self.i219156993,
      33689603: self.i33689603,
      167840769: self.i167840769,
      525572: self.i525572,
      134284806: self.i134284806
    }
    
  def i00(self): # nop
    pass

  def i01(self): # lit
    self.ip += 1
    self.data.lit(self.memory[self.ip])

  def i02(self): # dup
    self.data.raw_lit(self.data.top())

  def i03(self): # drop
    self.data.pop()

  def i04(self): # swap
    a = self.data.pop()
    b = self.data.pop()
    self.data.raw_lit(a)
    self.data.raw_lit(b)

  def i05(self): # push
    self.address.raw_lit(self.data.pop())

  def i06(self): # pop
    self.data.raw_lit(self.address.pop())

  def i07(self): # jump
    self.ip = self.data.pop() - 1

  def i08(self): # call
    self.address.raw_lit(self.ip)
    self.ip = self.data.pop() - 1

  def i09(self): # ccall
    target = self.data.pop()
    flag = self.data.pop()
    if flag != 0:
      self.address.raw_lit(self.ip)
      self.ip = target - 1

  def i10(self): # cjump
    target = self.data.pop()
    flag = self.data.pop()
    if flag != 0:
      self.ip = target - 1

  def i11(self): # return
    self.ip = self.address.pop()

  def i12(self): # equal
    self.data.raw_lit(-(self.data.pop() == self.data.pop()))

  def i13(self): # not equal
    self.data.raw_lit(-(self.data.pop() != self.data.pop()))

  def i14(self): # less than
    b = self.data.pop()
    a = self.data.pop()
    self.data.raw_lit(-(a < b))

  def i15(self): # greater than
    b = self.data.pop()
    a = self.data.pop()
    self.data.raw_lit(-(a > b))

  def i16(self): # fetch
    self.data.raw_lit(self.memory[self.data.pop()])

  def i17(self): # store
    a = self.data.pop()
    self.memory[a] = self.data.pop()

  def i18(self): # add
    self.data.lit(self.data.pop() + self.data.pop())

  def i19(self): # subtract
    b = self.data.pop()
    self.data.lit(self.data.pop() - b)
  
  def i20(self): # multiply
    self.data.lit(self.data.pop() * self.data.pop())
    
  def i21(self): # divmod
    a = self.data.pop()
    a, b = div_mod(self.data.pop(), a)
    self.data.raw_lit(b)
    self.data.raw_lit(a)

  def i22(self): # and
    self.data.raw_lit(self.data.pop() & self.data.pop())

  def i23(self): # or
    self.data.raw_lit(self.data.pop() | self.data.pop())

  def i24(self): # xor
    self.data.raw_lit(self.data.pop() ^ self.data.pop())

  def i25(self): # shift left
    b = self.data.pop()
    self.data.lit(self.data.pop() << b)

  def i26(self): # shift right
    b = self.data.pop()
    self.data.raw_lit(self.data.pop() >> b)

  def i27(self): # compare memory
    l = self.data.pop()
    dest = self.data.pop()
    src = self.data.pop()
    self.data.raw_lit(-(self.memory[dest: dest + l] == self.memory[src: src + l]))

  def i28(self): # copy memory
    l = self.data.pop()
    dest = self.data.pop()
    src = self.data.pop()
    self.memory[dest: dest + l] = self.memory[src: src + l]

  def i29(self): # io
    i = self.data.pop()
    if i == 0:
      print(chr(self.data.pop()), end='')
    elif i == 1:
      if not len(self.input_buffer):
        self.input_buffer = ['\n'] + list(input())[::-1]
      self.data.lit(ord(self.input_buffer.pop()))
    elif i == 2:
      self.block_tools.read_block()
    elif i == 3:
      self.block_tools.block_write()
    elif i == 4:
      self.block_tools.save_image()
    elif i == 5:
      self.block_tools.load_image()
      self.ip = -1
    elif i == 6:
      self.ip = 65535
    elif i == 7:
      self.data.lit(self.data.depth())
      self.data.lit(self.address.depth())

  def i2049(self):  # lica....
    self.address.lit(self.ip + 1)
    self.ip = self.memory[self.ip + 1] - 1

  def i1793(self):  # liju....
    self.ip = self.memory[self.ip + 1] - 1

  def i524545(self):  # lilica..
    self.data.lit(self.memory[self.ip + 1])
    self.address.lit(self.ip + 2)
    self.ip = self.memory[self.ip + 2] - 1

  def i459009(self):  # liliju..
    self.data.lit(self.memory[self.ip + 1])
    self.ip = self.memory[self.ip + 2] - 1

  def i67502597(self):  # puduposw 'over'
    self.data.lit(self.data.second())

  def i17563906(self):  # dulieqli
    self.data.lit(-(self.memory[self.ip + 1] == self.data.top()))
    self.data.lit(self.memory[self.ip + 2])
    self.ip += 2

  def i100926722(self):  # dupuswpo 'tuck'
    a = self.data.pop()
    b = self.data.pop()
    self.data.lit(a)
    self.data.lit(b)
    self.data.lit(a)

  def i302059522(self):  # dufeliad
    self.data.lit(self.memory[self.data.top()] + self.memory[self.ip + 1])
    self.ip += 1

  def i2818(self):  # dure....
    self.data.raw_lit(self.data.top())
    self.ip = self.address.pop()
  
  def i219156993(self):  # liadfene
    self.ip += 1
    self.data.lit(-(limit(self.memory[self.memory[self.ip] + self.data.pop()]) != self.data.pop()))
  
  def i33689603(self): # drfedudu
    self.data.pop()
    m = self.memory[self.data.pop()]
    self.data.lit(m)
    self.data.lit(m)
    self.data.lit(m)
  
  def i167840769(self):  # lieqlicj
    self.ip += 2
    if -(self.memory[self.ip-1] == self.data.pop()) != 0:
      self.ip = self.memory[self.ip] - 1
  
  def i2832(self): # fere....
    self.data.raw_lit(self.memory[self.data.pop()])
    self.ip = self.address.pop()
  
  def i68288770(self): #duliadsw
    self.data.raw_lit(self.data.top())
    self.ip += 1
    self.data.lit(limit(self.memory[self.ip]) + self.data.pop())
    a = self.data.pop()
    b = self.data.pop()
    self.data.raw_lit(a)
    self.data.raw_lit(b)
    
  def i525572(self): # swpuca..
    a = self.data.pop()
    self.address.raw_lit(self.data.pop())
    self.data.raw_lit(a)
    self.address.raw_lit(self.ip)
    self.ip = self.data.pop() - 1
  
  def i459014(self): # poliju..
    self.data.raw_lit(self.address.pop())
    self.ip = self.memory[self.ip+1] - 1

  def i134284806(self): # popolica
    self.data.raw_lit(self.address.pop())
    self.data.raw_lit(self.address.pop())
    self.ip += 1
    self.address.lit(self.ip)
    self.ip = self.memory[self.ip] - 1

  def print_opcode_bundle(self, opcode):
    print(opcode & 0xFF, end=' ')
    print((opcode >> 8) & 0xFF, end=' ')
    print((opcode >> 16) & 0xFF, end=' ')
    print((opcode >> 24) & 0xFF, end=': ')

  def process(self, c):
    if c in self.instructions:
      self.instructions[c]()

  def execute(self, start):
    self.ip = start
    while self.ip < 65536:
      m = self.memory[self.ip]
      if m in self.instructions:
        self.instructions[m]()
      else:
        self.process(m & 0xFF)
        self.process((m >> 8) & 0xFF)
        self.process((m >> 16) & 0xFF)
        self.process((m >> 24) & 0xFF)
      self.ip += 1
  
  def debug_execute(self, start):
    data = {}
    self.ip = start
    while self.ip < 65536:
      m = self.memory[self.ip]
      if m in self.instructions:
        self.instructions[m]()
      else:
        self.process(m & 0xFF)
        self.process((m >> 8) & 0xFF)
        self.process((m >> 16) & 0xFF)
        self.process((m >> 24) & 0xFF)
        data[str(m & 0xFF)] = data.get(str(m & 0xFF), 0) + 1
        data[str((m >> 8) & 0xFF)] = data.get(str((m >> 8) & 0xFF), 0) + 1
        data[str((m >> 16) & 0xFF)] = data.get(str((m >> 16) & 0xFF), 0) + 1
        data[str((m >> 24) & 0xFF)] = data.get(str((m >> 24) & 0xFF), 0) + 1
      data[str(m)] = data.get(str(m), 0) + 1
      self.ip += 1
    return data

  def main(self, debug=False):
    self.block_tools.load_image()
    if BLOCK_CACHE:
      self.block_tools.read_all_blocks()
    if debug:
      return self.debug_execute(0)
    else:
      self.execute(0)

def sort_dict(input_dict):
    sorted_items = sorted(input_dict.items(), key=lambda x: x[1], reverse=True)
    sorted_dict = dict(sorted_items)
    return sorted_dict

if __name__ == '__main__':
  if DEBUG:
    if DEBUG.lower() == 'instrs':
      import time
      start_time = time.time()
      k = ilo(BLOCK_FILE, ROM_FILE, BLOCK_CACHE)
      data = k.main(True)
      end_time = time.time()
      print(str(sort_dict(data)).replace(',', '\n').strip('{}'))
      print('\nElapsed Time: ', end_time - start_time)
    elif DEBUG.lower() == 'timing':
      import time
      start_time = time.time()
      ilo(BLOCK_FILE, ROM_FILE, BLOCK_CACHE).main()
      end_time = time.time()
      print('\nElapsed Time: ', end_time - start_time)
    elif DEBUG.lower() == 'cprofile':
      import cProfile
      cProfile.run('ilo(BLOCK_FILE, ROM_FILE, BLOCK_CACHE).main()', sort='tottime')
    else:
      print('Unknown debug mode')
  else:
    ilo(BLOCK_FILE, ROM_FILE, BLOCK_CACHE).main()