#include "vm.h"
#include "device.h"
#include "opcodes.h"
#include "str.h"

#define COMPARE_AND_JUMP(type, op)                                             \
  do {                                                                         \
    i32 cond;                                                                  \
    u32 mask, target;                                                          \
    u8 src1, src2;                                                             \
    type value;                                                                \
    type value2;                                                               \
    target = read_u32(vm, code, vm->pc);                                       \
    vm->pc += 4;                                                               \
    src1 = read_u8(vm, code, vm->pc);                                          \
    vm->pc++;                                                                  \
    src2 = read_u8(vm, code, vm->pc);                                          \
    vm->pc++;                                                                  \
    value = (type)frame->registers[src1];                                      \
    value2 = (type)frame->registers[src2];                                     \
    cond = !!(value op value2);                                                \
    mask = -(u32)cond;                                                         \
    vm->pc = (target & mask) | (vm->pc & ~mask);                               \
    return true;                                                               \
  } while (0)

#define MATH_OP(type, op)                                                      \
  do {                                                                         \
    dest = read_u8(vm, code, vm->pc);                                          \
    vm->pc++;                                                                  \
    src1 = read_u8(vm, code, vm->pc);                                          \
    vm->pc++;                                                                  \
    src2 = read_u8(vm, code, vm->pc);                                          \
    vm->pc++;                                                                  \
    frame->registers[dest] =                                                   \
        (type)frame->registers[src1] op(type) frame->registers[src2];          \
    return true;                                                               \
  } while (0)

#define BIT_OP(op)                                                             \
  do {                                                                         \
    dest = read_u8(vm, code, vm->pc);                                          \
    vm->pc++;                                                                  \
    src1 = read_u8(vm, code, vm->pc);                                          \
    vm->pc++;                                                                  \
    src2 = read_u8(vm, code, vm->pc);                                          \
    vm->pc++;                                                                  \
    frame->registers[dest] = frame->registers[src1] op frame->registers[src2]; \
    return true;                                                               \
  } while (0)

u32 str_alloc(VM *vm, Frame *frame, const char *str, u32 length) {
  u32 str_addr = vm->mp;
  u32 i = 0;
  vm->mp += 4;
  while (i < length) {
    vm->memory[vm->mp++] = str[i++];
  }
  vm->memory[vm->mp++] = '\0';

  write_u32(vm, memory, str_addr, length);
  frame->end = vm->mp;
  return str_addr;
}

/**
 * Step to the next opcode in the vm.
 */
bool step_vm(VM *vm) {
  u16 opcode, dest, src1, src2;
  u32 v, ptr;
  i32 value;
  Frame *frame;

  /* Get current instruction & Advance to next instruction */
  opcode = vm->code[vm->pc++];
  frame = &vm->frames[vm->fp];

  switch (opcode) {
  case OP_HALT: {
    return false;
  }
  case OP_CALL: {
    u32 jmp = read_u32(vm, code, vm->pc); /* location of function in code */
    vm->pc += 4;
    vm->return_stack[vm->rp++] = vm->pc; /* set return address */
    vm->fp++;                            /* increment to the next free frame */
    vm->frames[vm->fp].start = vm->mp;   /* set start of new memory block */
    vm->pc = jmp;
    return true;
  }
  case OP_RETURN: {
    frame->rp = 0;                       /* reset register ptr */
    vm->pc = vm->return_stack[--vm->rp]; /* set pc to return address */
    vm->mp =
        vm->frames[vm->fp--].start; /* reset memory pointer to start
                                                 of old slice, pop the frame */
    return true;
  }
  case OP_MALLOC: {
    u32 size;
    dest = read_u8(vm, code, vm->pc);
    vm->pc++;
    src1 = read_u8(vm, code, vm->pc);
    vm->pc++;
    frame->registers[dest] = vm->mp;
    size = frame->registers[src1];
    write_u32(vm, memory, vm->mp, size);
    vm->mp += (size + 4);
    return true;
  }
  case OP_MEMSET_32: {
    u32 i, start, end;
    u8 dest_reg = read_u8(vm, code, vm->pc++);
    u8 value_reg = read_u8(vm, code, vm->pc++);
    u8 count_reg = read_u8(vm, code, vm->pc++);
    
    u32 dest = frame->registers[dest_reg];
    u32 value = frame->registers[value_reg];
    u32 count = frame->registers[count_reg];

    if (count == 0) {
      vm->flag = 1;
      return true;
    }

    start = dest;
    end = dest + count;

    if (start >= vm->mp || count > vm->mp ||
        end > vm->mp) {
      vm->flag = 0;
      return true;
    }

    for (i = start; i < end; i+=4) {
      write_u32(vm, memory, i, value);
    }

    frame->registers[0] = dest;
    vm->flag = 1;
    return true;
  }
  case OP_MEMSET_16: {
    u32 i, start, end;
    u8 dest_reg = read_u8(vm, code, vm->pc++);
    u8 value_reg = read_u8(vm, code, vm->pc++);
    u8 count_reg = read_u8(vm, code, vm->pc++);
    
    u32 dest = frame->registers[dest_reg];
    u16 value = (u16)(frame->registers[value_reg]);
    u32 count = frame->registers[count_reg];

    if (count == 0) {
      vm->flag = 1;
      return true;
    }

    start = dest;
    end = dest + count;

    if (start >= vm->mp || count > vm->mp ||
        end > vm->mp) {
      vm->flag = 0;
      return true;
    }

    for (i = start; i < end; i+=2) {
      write_u16(vm, memory, i, value);
    }

    frame->registers[0] = dest;
    vm->flag = 1;
    return true;
  }
  case OP_MEMSET_8: {
    u32 i, start, end;
    u8 dest_reg = read_u8(vm, code, vm->pc++);
    u8 value_reg = read_u8(vm, code, vm->pc++);
    u8 count_reg = read_u8(vm, code, vm->pc++);
    
    u32 dest = frame->registers[dest_reg];
    u8 value = (u8)(frame->registers[value_reg]);
    u32 count = frame->registers[count_reg];

    if (count == 0) {
      vm->flag = 1;
      return true;
    }

    start = dest;
    end = dest + count;

    if (start >= vm->mp || count > vm->mp ||
        end > vm->mp) {
      vm->flag = 0;
      return true;
    }

    for (i = start; i < end; i++) {
      write_u8(vm, memory, i, value);
    }

    frame->registers[0] = dest;
    vm->flag = 1;
    return true;
  }
  case OP_LOAD_IMM: {
    dest = read_u8(vm, code, vm->pc);
    vm->pc++;
    v = read_u32(vm, code, vm->pc);
    vm->pc += 4;
    frame->registers[dest] = v;
    return true;
  }
  case OP_LOAD_32: {
    dest = read_u8(vm, code, vm->pc);
    vm->pc++;
    ptr = read_u32(vm, code, vm->pc);
    vm->pc += 4;
    v = read_u32(vm, memory, ptr);
    frame->registers[dest] = v;
    return true;
  }
  case OP_LOAD_16: {
    dest = read_u8(vm, code, vm->pc);
    vm->pc++;
    ptr = read_u32(vm, code, vm->pc);
    vm->pc += 4;
    v = read_u16(vm, memory, ptr);
    frame->registers[dest] = v;
    return true;
  }
  case OP_LOAD_8: {
    dest = read_u8(vm, code, vm->pc);
    vm->pc++;
    ptr = read_u32(vm, code, vm->pc);
    vm->pc += 4;
    v = read_u8(vm, memory, ptr);
    frame->registers[dest] = v;
    return true;
  }
  case OP_GET_32: {
    dest = read_u8(vm, code, vm->pc);
    vm->pc++;
    src1 = read_u8(vm, code, vm->pc);
    vm->pc++;
    v = frame->registers[src1];
    ptr = read_u32(vm, memory, v);
    frame->registers[dest] = ptr;
    return true;
  }
  case OP_GET_16: {
    u16 v16;
    dest = read_u8(vm, code, vm->pc);
    vm->pc++;
    src1 = read_u8(vm, code, vm->pc);
    vm->pc++;
    v = frame->registers[src1];
    v16 = read_u16(vm, memory, v);
    frame->registers[dest] = v16;
    return true;
  }
  case OP_GET_8: {
    u8 v8;
    dest = read_u8(vm, code, vm->pc);
    vm->pc++;
    src1 = read_u8(vm, code, vm->pc);
    vm->pc++;
    v = frame->registers[src1];
    v8 = read_u8(vm, memory, v);
    frame->registers[dest] = v8;
    return true;
  }
  case OP_STORE_32: {
    dest = read_u8(vm, code, vm->pc);
    vm->pc++;
    src1 = read_u8(vm, code, vm->pc);
    vm->pc++;
    v = frame->registers[src1];
    ptr = frame->registers[dest];
    write_u32(vm, memory, ptr, v);
    return true;
  }
  case OP_STORE_16: {
    dest = read_u8(vm, code, vm->pc);
    vm->pc++;
    src1 = read_u8(vm, code, vm->pc);
    vm->pc++;
    v = frame->registers[src1];
    ptr = frame->registers[dest];
    write_u16(vm, memory, ptr, v);
    return true;
  }
  case OP_STORE_8: {
    dest = read_u8(vm, code, vm->pc);
    vm->pc++;
    src1 = read_u8(vm, code, vm->pc);
    vm->pc++;
    v = frame->registers[src1];
    ptr = frame->registers[dest];
    write_u8(vm, memory, ptr, v);
    return true;
  }
  case OP_PUT_32: {
    dest = read_u8(vm, code, vm->pc);
    vm->pc++;
    src1 = read_u8(vm, code, vm->pc);
    vm->pc++;
    ptr = frame->registers[dest];
    v = frame->registers[src1];    
    write_u32(vm, memory, ptr, v);    
    return true;
  }
  case OP_PUT_16: {
    u16 v16;
    dest = read_u8(vm, code, vm->pc);
    vm->pc++;
    src1 = read_u8(vm, code, vm->pc);
    vm->pc++;
    ptr = frame->registers[dest];
    v16 = frame->registers[src1];
    write_u16(vm, memory, ptr, v16);
    return true;
  }
  case OP_PUT_8: {
    u8 v8;
    dest = read_u8(vm, code, vm->pc);
    vm->pc++;
    src1 = read_u8(vm, code, vm->pc);
    vm->pc++;
    ptr = frame->registers[dest];
    v8 = frame->registers[src1];
    write_u8(vm, memory, ptr, v8);
    return true;
  }
  case OP_PUSH: {
    dest = read_u8(vm, code, vm->pc);
    vm->pc++;
    vm->stack[++vm->sp] = frame->registers[dest];
    return true;
  }
  case OP_POP: {
    dest = read_u8(vm, code, vm->pc);
    vm->pc++;
    frame->registers[dest] = vm->stack[vm->sp--];
    return true;
  }
  case OP_REG_MOV: {
    dest = read_u8(vm, code, vm->pc);
    vm->pc++;
    src1 = read_u8(vm, code, vm->pc);
    vm->pc++;
    frame->registers[dest] = frame->registers[src1];
    return true;
  }
  case OP_JMP: {
    dest = read_u8(vm, code, vm->pc);
    vm->pc++;
    vm->pc = frame->registers[dest]; /* Jump to address */
    return true;
  }
  case OP_JMPF: { /* error handling for syscall, jump if flag == 0 */
    u32 mask;
    dest = read_u8(vm, code, vm->pc);
    vm->pc++;
    mask = -(u32)(vm->flag == 0);
    vm->pc = (dest & mask) | (vm->pc & ~mask);
    return true;
  }
  case OP_SYSCALL: {
    u32 syscall_id;

    syscall_id = read_u32(vm, code, vm->pc);
    vm->pc += 4;

    switch (syscall_id) {
    case SYSCALL_DEVICE_OPEN: {
      Device *dev;
      u32 path_ptr, mode;
      u8 path_reg, mode_reg, dest_reg;
      dest_reg = read_u8(vm, code, vm->pc);
      vm->pc++;
      path_reg = read_u8(vm, code, vm->pc);
      vm->pc++;
      mode_reg = read_u8(vm, code, vm->pc);
      vm->pc++;

      path_ptr = frame->registers[path_reg];
      mode = frame->registers[mode_reg];
      dev = find_device_by_path(vm, (const char *)&vm->memory[path_ptr + 4]);
      if (dev) {
        if (dev->ops->open) {
          vm->flag = dev->ops->open(dev->data, mode);
          frame->registers[dest_reg] = dev->handle;
        } else {
          vm->flag = 1; /* success, no open needed */
        }
      } else {
        vm->flag = 0; /* error */
      }

      return true;
    }

    case SYSCALL_DEVICE_READ: {
      Device *dev;
      u32 handle, buffer_ptr, size;
      u16 handle_reg, buffer_reg, size_reg;
      handle_reg = read_u8(vm, code, vm->pc);
      vm->pc++;
      buffer_reg = read_u8(vm, code, vm->pc);
      vm->pc++;
      size_reg = read_u8(vm, code, vm->pc);
      vm->pc++;
      dest = read_u8(vm, code, vm->pc);
      vm->pc++;

      handle = frame->registers[handle_reg]; /* path pointer */
      size = frame->registers[size_reg];     /* size */
      buffer_ptr = frame->registers[dest];

      dev = &vm->devices[handle];
      if (dev && dev->ops->read) {
        vm->flag = dev->ops->read(dev->data, &vm->memory[buffer_ptr + 4], size);
        frame->registers[buffer_reg] = buffer_ptr;
      } else {
        vm->flag = 0;
      }

      return true;
    }

    case SYSCALL_DEVICE_WRITE: {
      Device *dev;
      u32 handle, buffer_ptr, size;
      u16 handle_reg, buffer_reg, size_reg;
      handle_reg = read_u8(vm, code, vm->pc);
      vm->pc++;
      buffer_reg = read_u8(vm, code, vm->pc);
      vm->pc++;
      size_reg = read_u8(vm, code, vm->pc);
      vm->pc++;

      handle = frame->registers[handle_reg];     /* R0: path pointer */
      buffer_ptr = frame->registers[buffer_reg]; /* R1: buffer pointer */
      size = frame->registers[size_reg];         /* R2: size */

      dev = &vm->devices[handle];
      if (dev && dev->ops->write) {
        vm->flag = dev->ops->write(
            dev->data, (const u8 *)&vm->memory[buffer_ptr + 4], size);
      } else {
        vm->flag = 0;
      }

      return true;
    }

    case SYSCALL_DEVICE_CLOSE: {
      Device *dev;
      u32 handle;
      u8 handle_reg;
      handle_reg = read_u8(vm, code, vm->pc);
      vm->pc++;

      handle = frame->registers[handle_reg]; /* R0: path pointer */

      dev = &vm->devices[handle];
      if (dev && dev->ops->close) {
        i32 result = dev->ops->close(dev->data);
        vm->flag = result;
      } else {
        vm->flag = 0;
      }

      return true;
    }

    case SYSCALL_DEVICE_IOCTL: {
      Device *dev;
      u32 handle, args_ptr, cmd;
      u8 handle_reg, cmd_reg, args_ptr_reg;
      handle_reg = read_u8(vm, code, vm->pc);
      vm->pc++;
      cmd_reg = read_u8(vm, code, vm->pc);
      vm->pc++;
      args_ptr_reg = read_u8(vm, code, vm->pc);
      vm->pc++;

      handle = frame->registers[handle_reg];     /* R0: device path */
      cmd = frame->registers[cmd_reg];           /* R1: ioctl command */
      args_ptr = frame->registers[args_ptr_reg]; /* R2: args pointer */

      dev = &vm->devices[handle];
      if (dev && dev->ops && dev->ops->ioctl) {
        i32 result = dev->ops->ioctl(dev->data, cmd, &vm->memory[args_ptr]);
        vm->flag = result;
      } else {
        vm->flag = 0; /* error or no ioctl support */
      }

      return true;
    }

    case SYSCALL_EXIT: {
      return false;
    }

    default: {
      vm->flag = 0; /* unknown syscall */
      return true;
    }
    }
    return true;
  }
  case OP_SLL:
    BIT_OP(<<);
  case OP_SRL:
    BIT_OP(>>);
  case OP_SRE:
    MATH_OP(i32, >>);
  case OP_BAND:
    BIT_OP(&);
  case OP_BOR:
    BIT_OP(|);
  case OP_BXOR:
    BIT_OP(^);
  case OP_ADD_INT:
    MATH_OP(i32, +);
  case OP_SUB_INT:
    MATH_OP(i32, -);
  case OP_MUL_INT:
    MATH_OP(i32, *);
  case OP_DIV_INT:
    MATH_OP(i32, /);
  case OP_ADD_UINT:
    MATH_OP(u32, +);
  case OP_SUB_UINT:
    MATH_OP(u32, -);
  case OP_MUL_UINT:
    MATH_OP(u32, *);
  case OP_DIV_UINT:
    MATH_OP(u32, /);
  case OP_MUL_REAL: {
    dest = read_u8(vm, code, vm->pc);
    vm->pc++;
    src1 = read_u8(vm, code, vm->pc);
    vm->pc++;
    src2 = read_u8(vm, code, vm->pc);
    vm->pc++;
    frame->registers[dest] =
        (frame->registers[src1] * frame->registers[src2]) >> 16;
    return true;
  }

  case OP_DIV_REAL: {
    dest = read_u8(vm, code, vm->pc);
    vm->pc++;
    src1 = read_u8(vm, code, vm->pc);
    vm->pc++;
    src2 = read_u8(vm, code, vm->pc);
    vm->pc++;
    frame->registers[dest] =
        (frame->registers[src1] << 16) / frame->registers[src2];
    return true;
  }

  case OP_ADD_REAL: {
    dest = read_u8(vm, code, vm->pc);
    vm->pc++;
    src1 = read_u8(vm, code, vm->pc);
    vm->pc++;
    src2 = read_u8(vm, code, vm->pc);
    vm->pc++;
    frame->registers[dest] = frame->registers[src1] + frame->registers[src2];
    return true;
  }

  case OP_SUB_REAL: {
    dest = read_u8(vm, code, vm->pc);
    vm->pc++;
    src1 = read_u8(vm, code, vm->pc);
    vm->pc++;
    src2 = read_u8(vm, code, vm->pc);
    vm->pc++;
    frame->registers[dest] = frame->registers[src1] - frame->registers[src2];
    return true;
  }
  case OP_REAL_TO_INT: {
    dest = read_u8(vm, code, vm->pc);
    vm->pc++;
    src1 = read_u8(vm, code, vm->pc);
    vm->pc++;
    value = frame->registers[src1];

    if (value >= 0) {
      frame->registers[dest] = value >> 16;
    } else {
      frame->registers[dest] = -((-value) >> 16);
    }

    return true;
  }
  case OP_INT_TO_REAL: {
    dest = read_u8(vm, code, vm->pc);
    vm->pc++;
    src1 = read_u8(vm, code, vm->pc);
    vm->pc++;
    frame->registers[dest] = (frame->registers[src1] << 16);
    return true;
  }
  case OP_REAL_TO_UINT: {
    dest = read_u8(vm, code, vm->pc);
    vm->pc++;
    src1 = read_u8(vm, code, vm->pc);
    vm->pc++;
    value = frame->registers[src1];
    if (value < 0) {
      frame->registers[dest] = 0;
    } else {
      frame->registers[dest] = AS_UINT(value >> 16);
    }
    return true;
  }
  case OP_UINT_TO_REAL: {
    dest = read_u8(vm, code, vm->pc);
    vm->pc++;
    src1 = read_u8(vm, code, vm->pc);
    vm->pc++;
    frame->registers[dest] = AS_INT(frame->registers[src1] << 16);
    return true;
  }
  case OP_JEQ_UINT: {
    COMPARE_AND_JUMP(u32, ==);
  }
  case OP_JNEQ_UINT: {
    COMPARE_AND_JUMP(u32, !=);
  }
  case OP_JGT_UINT: {
    COMPARE_AND_JUMP(u32, >);
  }
  case OP_JLT_UINT: {
    COMPARE_AND_JUMP(u32, <);
  }
  case OP_JLE_UINT: {
    COMPARE_AND_JUMP(u32, <=);
  }
  case OP_JGE_UINT: {
    COMPARE_AND_JUMP(u32, >=);
  }
  case OP_JEQ_INT: {
    COMPARE_AND_JUMP(i32, ==);
  }
  case OP_JNEQ_INT: {
    COMPARE_AND_JUMP(i32, !=);
  }
  case OP_JGT_INT: {
    COMPARE_AND_JUMP(i32, >);
  }
  case OP_JLT_INT: {
    COMPARE_AND_JUMP(i32, <);
  }
  case OP_JLE_INT: {
    COMPARE_AND_JUMP(i32, <=);
  }
  case OP_JGE_INT: {
    COMPARE_AND_JUMP(i32, >=);
  }
  case OP_JEQ_REAL: {
    COMPARE_AND_JUMP(i32, ==);
  }
  case OP_JNEQ_REAL: {
    COMPARE_AND_JUMP(i32, !=);
  }
  case OP_JGT_REAL: {
    COMPARE_AND_JUMP(i32, >);
  }
  case OP_JLT_REAL: {
    COMPARE_AND_JUMP(i32, <);
  }
  case OP_JGE_REAL: {
    COMPARE_AND_JUMP(i32, >=);
  }
  case OP_JLE_REAL: {
    COMPARE_AND_JUMP(i32, <=);
  }
  case OP_INT_TO_STRING: {
    char buffer[32];
    dest = read_u8(vm, code, vm->pc);
    vm->pc++;
    src1 = read_u8(vm, code, vm->pc);
    vm->pc++;
    int_to_string(AS_INT(frame->registers[src1]), buffer);
    ptr = str_alloc(vm, frame, buffer, strlength(buffer));
    frame->registers[dest] = ptr;
    return true;
  }
  case OP_UINT_TO_STRING: {
    char buffer[32];
    dest = read_u8(vm, code, vm->pc);
    vm->pc++;
    src1 = read_u8(vm, code, vm->pc);
    vm->pc++;
    uint_to_string(frame->registers[src1], buffer);
    ptr = str_alloc(vm, frame, buffer, strlength(buffer));
    frame->registers[dest] = ptr;
    return true;
  }
  case OP_REAL_TO_STRING: {
    char buffer[32];
    dest = read_u8(vm, code, vm->pc);
    vm->pc++;
    src1 = read_u8(vm, code, vm->pc);
    vm->pc++;
    fixed_to_string(AS_INT(frame->registers[src1]), buffer);
    ptr = str_alloc(vm, frame, buffer,
                    strlength(buffer)); /* copy buffer to dest */
    frame->registers[dest] = ptr;
    return true;
  }
  case OP_STRLEN: {
    u32 ptr, length;
    dest = read_u8(vm, code, vm->pc);
    vm->pc++;
    src1 = read_u8(vm, code, vm->pc);
    vm->pc++;

    ptr = frame->registers[src1];
    length = read_u32(vm, memory, ptr);
    frame->registers[dest] = length;
    return true;
  }
  case OP_STRCAT: {
    /* not implemented yet */
    return false;
  }
  case OP_STR_GET_CHAR: {
    /* not implemented yet */
    return false;
  }
  case OP_STR_FIND_CHAR: {
    /* not implemented yet */
    return false;
  }
  case OP_STR_SLICE: {
    /* not implemented yet */
    return false;
  }
  case OP_STRING_TO_INT: {
    /* not implemented yet */
    return false;
  }
  case OP_STRING_TO_UINT: {
    /* not implemented yet */
    return false;
  }
  case OP_STRING_TO_REAL: {
    /* not implemented yet */
    return false;
  }
  }
  return false; /* something bad happened */
}