#include "vm.h"
#define FRAME_HEADER_SIZE 12

u32 pc;       /* program counter */
u32 cp;       /* code pointer */
u32 mp;       /* memory pointer */
u32 fp;       /* frame pointer */
u8  lc;       /* child local count */
u8  status;   /* status flag */
u8 interrupt; /* device interrupt */
u32 *code;    /* code */
u8 *mem;      /* memory */

#define MAX_LEN_INT32 11
#define MAX_INT32 2147483647
#define MIN_INT32 -2147483648
const char radix_set[11] = "0123456789";

u32 str_alloc(char *str, u32 length) {
  u32 str_addr = mp;
  u8 *ptr = &mem[mp];
  mcpy(ptr, &length, sizeof(u32));
  ptr += 4;
  mcpy(ptr, str, length);
  ptr[length] = '\0';
  mp += 4 + length + 1;
  return str_addr;
}

bool step_vm() {

  u32 instruction = code[pc++];
  u8 opcode = DECODE_OP(instruction);
  u32 *locals = (u32*)(&mem[fp]);
  u32 *globals = (u32*)(mem);

  switch (opcode) {
  case OP_HALT: {
    /* no need to decode, all are zeros */
    return false;
  }
  case OP_CALL: {
    DECODE_A(instruction)
    /* function to jump to */
    u32 fn_ptr = locals[dest]; 
    /* get mp in 'global indexing mode' */
    u32 *header = &globals[mp / 4];
    /* reset child locals counter */
    lc = 0;
    /* push parents frame value to reset the heap to */
    (*header++) = fp;
    /* push return address to child frame */
    (*header++) = pc;
    /* push local address to return the value to */
    (*header++) = fp + src2;
    /* increase the mp to new size */
    mp += FRAME_HEADER_SIZE;
    /* now set the frame pointer, where the locals start */
    fp = mp;    
    /* move mp forward by count many locals */
    mp += src1;
    /* jump to dest_ptr */
    pc = fn_ptr;
    return true;
  }
  case OP_RETURN: {
    DECODE_B(instruction)
    u32 size = 0;
    u32 return_value = locals[dest];
    bool is_ptr = (((u32)(1)) << 15) & imm;
    bool replaces_value = (((u32)(1)) << 14) & imm;

    /* reset mp to saved mp, use header size to get "real" start of frame */
    u32 *frame_start = &globals[(fp / 4) - 3];
    u32 parent_fp = *frame_start++;
    u32 return_address = *frame_start++;
    u32 parent_local_return_address = *frame_start++;

    USED(replaces_value);
    /* reset memory to parents end of memory */
    mp = fp - FRAME_HEADER_SIZE;
    /* reset the frame pointer */
    fp = parent_fp;

    if (is_ptr) {
    /* copy value to end of mp if it is a pointer */
      globals[parent_local_return_address/4] = mp;
      size = globals[return_value/4];
      globals[mp/4] = size;
      mp += 4;
      mcpy(&mem[mp], &mem[return_value], size);
      mp += size;
    } else {
    /* otherwise just write the return value to its location */
      mcpy(&mem[parent_local_return_address], &return_value, sizeof(u32));
    }

    /* jump to parent frame */
    pc = return_address;
    return true;
  }
  case OP_SYSCALL: {
    DECODE_A(instruction)
    u32 id = dest;   /* syscall id */
    u32 size = src1; /* size of heap at that pointer */
    u32 rd = fp + (src2 * 4); /* the pointer */
    status = syscall(id, size, rd);
    return true;
  }
  case OP_PUSH: {
    DECODE_B(instruction)
    USED(imm); 
    globals[(mp / 4) + lc + 3] = locals[dest];
    lc++;
    return true;
  }
  case OP_POP: {
    DECODE_B(instruction)
    USED(dest);
    USED(imm); 
    mp -= 4;
    lc--;
    return true;
  }
  case OP_LOAD_IMM: {
    DECODE_B(instruction)
    locals[dest] = imm;
    return true;
  }
  case OP_LOAD_UPPER_IMM: {
    DECODE_B(instruction)
    u32 value = locals[dest];
    locals[dest] = (value | (((u32)(imm)) << 16));
    return true;
  }
  case OP_LOAD_IND_8: {
  }
  case OP_LOAD_IND_16: {
  }
  case OP_LOAD_IND_32: {
  }
  case OP_LOAD_ABS_8: {
  }
  case OP_LOAD_ABS_16: {
  }
  case OP_LOAD_ABS_32: {
  }
  case OP_LOAD_OFF_8: {
  }
  case OP_LOAD_OFF_16: {
  }
  case OP_LOAD_OFF_32: {
  }
  case OP_STORE_ABS_8: {
  }
  case OP_STORE_ABS_16: {
  }
  case OP_STORE_ABS_32: {
  }
  case OP_STORE_IND_8: {
  }
  case OP_STORE_IND_16: {
  }
  case OP_STORE_IND_32: {
  }
  case OP_STORE_OFF_8: {
  }
  case OP_STORE_OFF_16: {
  }
  case OP_STORE_OFF_32: {
  }
  case OP_MEM_ALLOC: {
    DECODE_A(instruction)
    u32 size = locals[src1];
    locals[dest] = mp;
    WRITE_U32(mp, size);
    USED(src2);
    mp += (size + 4);
    return true;
  }
  case OP_MEM_CPY_8: {
    DECODE_A(instruction)
    u32 i = 0;
    u32 mdest = locals[dest];
    u32 msrc = locals[src1];
    u32 count = locals[src2];

    if (mdest + count >= mp) {
      status = 1;
      return true;
    }

    for (i = 0; i < count; i++) {
      mem[msrc + i] = mem[mdest + i];
    }

    status = 0;
    return true;
  }
  case OP_MEM_CPY_16: {
    DECODE_A(instruction)
    u32 i = 0;
    u32 mdest = locals[dest];
    u32 msrc = locals[src1];
    u32 count = locals[src2] * 2;

    if (mdest + count >= mp) {
      status = 1;
      return true;
    }

    for (i = 0; i < count; i++) {
      u16 value = READ_U16(mdest + i);
      WRITE_U16(msrc + i, value);
    }

    status = 0;
    return true;
  }
  case OP_MEM_CPY_32: {
    DECODE_A(instruction)
    u32 i = 0;
    u32 mdest = locals[dest];
    u32 msrc = locals[src1];
    u32 count = locals[src2];

    if (mdest + count >= mp) {
      status = 1;
      return true;
    }

    for (i = 0; i < count; i++) {
      globals[msrc + i] = globals[mdest + i];
    }

    status = 0;
    return true;
  }
  case OP_MEM_SET_8: {
    DECODE_A(instruction)
    u32 i, start, end;

    u32 rd = fp + (dest * 4);
    u32 r1 = fp + (src1 * 4);
    u32 r2 = fp + (src2 * 4);

    u8 value = (u8)READ_U32(r1);
    u32 count = READ_U32(r2);

    if (r2 == 0) {
      status = 1;
      return true;
    }

    start = READ_U32(rd);
    end = start + count;

    if (start >= mp || r2 > mp || end > mp) {
      status = 1;
      return true;
    }

    for (i = start; i < end; i++) {
      mem[i] = value;
    }

    status = 0;
    return true;
  }
  case OP_MEM_SET_16: {
    DECODE_A(instruction)
    u32 i, start, end;

    u32 rd = fp + (dest * 4);
    u32 r1 = fp + (src1 * 4);
    u32 r2 = fp + (src2 * 4);

    u16 value = (u16)READ_U32(r1);
    u32 count = READ_U32(r2);

    if (r2 == 0) {
      status = 1;
      return true;
    }

    start = READ_U32(rd);
    end = start + count;

    if (start >= mp || r2 > mp || end > mp) {
      status = 1;
      return true;
    }

    for (i = start; i < end; i += 2) {
      WRITE_U16(i, value);
    }

    status = 0;
    return true;
  }
  case OP_MEM_SET_32: {
    DECODE_A(instruction)
    u32 i, start, end;

    u32 value = locals[src1];
    u32 count = locals[src2];

    if (count == 0) {
      status = 1;
      return true;
    }

    start = READ_U32(locals[dest]);
    end = start + count;

    if (start >= mp || count > mp || end > mp) {
      status = 1;
      return true;
    }

    for (i = start; i < end; i += 4) {
      WRITE_U32(i, value);
    }

    status = 0;
    return true;
  }
  case OP_MOV: {
    DECODE_A(instruction)
    USED(src2);
    locals[dest] = locals[src1];
    return true;
  }
  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_NAT: {
    MATH_OP(u32, +);
  }
  case OP_SUB_NAT: {
    MATH_OP(u32, -);
  }
  case OP_MUL_NAT: {
    MATH_OP(u32, *);
  }
  case OP_DIV_NAT: {
    MATH_OP(u32, /);
  }
  case OP_ADD_REAL: {
    MATH_OP(i32, +);
  }
  case OP_SUB_REAL: {
    MATH_OP(i32, -);
  }
  case OP_MUL_REAL: {
    DECODE_A(instruction)

    i32 src1_whole = (i32)locals[src1] >> 16;
    i32 src2_whole = (i32)locals[src2] >> 16;

    i32 src1_decimal = (i32)locals[src1] & 16;
    i32 src2_decimal = (i32)locals[src2] & 16;

    i32 result = 0;
    result += (src1_whole * src2_whole) << 16;
    result += (src1_whole * src2_decimal);
    result += (src1_decimal * src2_whole);
    result += ((src1_decimal * src2_decimal) >> 16) & 16;
    locals[dest] = result;
    return true;
  }
  case OP_DIV_REAL: {
    DECODE_A(instruction)
    i32 result;
    i32 src1_val = (i32)locals[src1];
    i32 src2_val = (i32)locals[src2];

    u32 src2_reciprocal = 1;
    src2_reciprocal <<= 31;
    src2_reciprocal = (u32)(src2_reciprocal / src2_val);

    result = src1_val * src2_reciprocal;
    result <<= 1;

    locals[dest] = result;
    return true;
  }
  case OP_INT_TO_REAL: {
    DECODE_A(instruction)
    i32 result = (i32)locals[src1] << 16;
    USED(src2);
    locals[dest] = result;
    return true;
  }
  case OP_INT_TO_NAT: {
    DECODE_A(instruction)
    u32 result = (u32)locals[src1];
    USED(src2);
    locals[dest] = result;
    return true;
  }
  case OP_NAT_TO_REAL: {
    DECODE_A(instruction)
    i32 result = (i32)locals[src1] << 16;
    USED(src2);
    locals[dest] = result;
    return true;
  }
  case OP_NAT_TO_INT: {
    DECODE_A(instruction)
    i32 result = (i32)locals[src1];
    USED(src2);
    locals[dest] = result;
    return true;
  }
  case OP_REAL_TO_INT: {
    DECODE_A(instruction)
    i32 result = (i32)locals[src1] >> 16;
    USED(src2);
    locals[dest] = result;
    return true;
  }
  case OP_REAL_TO_NAT: {
    DECODE_A(instruction)
    u32 result = (u32)locals[src1] >> 16;
    USED(src2);
    locals[dest] = result;
    return true;
  }
  case OP_BIT_SHIFT_LEFT: {
    MATH_OP_NO_CAST(<<);
  }
  case OP_BIT_SHIFT_RIGHT: {
    MATH_OP_NO_CAST(>>);
  }
  case OP_BIT_SHIFT_R_EXT: {
    MATH_OP(i32, >>);
  }
  case OP_BIT_AND: {
    MATH_OP_NO_CAST(&);
  }
  case OP_BIT_OR: {
    MATH_OP_NO_CAST(|);
  }
  case OP_BIT_XOR: {
    MATH_OP_NO_CAST(^);
  }
  case OP_JMP_IMM: {
   u32 imm = (((u32)code[(pc) + 3] << 24) |       
      ((u32)code[(pc) + 2] << 16) |          
      ((u32)code[(pc) + 1] << 8) | 
      ((u32)code[(pc)]));
    pc = imm;
    return true;
  }
  case OP_JMP_ABS: {
    DECODE_A(instruction)
    u32 jmp_dest = locals[dest];
    if (jmp_dest > cp) {
      status = 1;
      return true;
    }
    USED(src1);
    USED(src2);

    pc = jmp_dest;
    return true;
  }
  case OP_JMP_OFF: {
    DECODE_A(instruction)
    u32 jmp_dest = locals[dest] + locals[src1];
    if (jmp_dest > cp) {
      status = 1;
      return true;
    }
    USED(src2);

    pc = jmp_dest;
    return true;
  }
  case OP_JMP_FLAG: {
    DECODE_A(instruction)
    u32 mask;
    u32 jmp_dest = locals[dest];
    if (jmp_dest > cp) {
      status = 1;
      return true;
    }
    USED(src1);
    USED(src2);

    mask = -(u32)(status == 0);
    pc = (jmp_dest & mask) | (pc & ~mask);
    return true;
  }
  case OP_JEQ_INT: {
    COMPARE_AND_JUMP(i32, ==);
  }
  case OP_JNE_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_NAT: {
    COMPARE_AND_JUMP(u32, ==);
  }
  case OP_JNE_NAT: {
    COMPARE_AND_JUMP(u32, !=);
  }
  case OP_JGT_NAT: {
    COMPARE_AND_JUMP(u32, >);
  }
  case OP_JLT_NAT: {
    COMPARE_AND_JUMP(u32, <);
  }
  case OP_JLE_NAT: {
    COMPARE_AND_JUMP(u32, <=);
  }
  case OP_JGE_NAT: {
    COMPARE_AND_JUMP(u32, >=);
  }
  case OP_JEQ_REAL: {
    COMPARE_AND_JUMP(i32, ==);
  }
  case OP_JNE_REAL: {
    COMPARE_AND_JUMP(i32, !=);
  }
  case OP_JGE_REAL: {
    COMPARE_AND_JUMP(i32, >=);
  }
  case OP_JGT_REAL: {
    COMPARE_AND_JUMP(i32, >);
  }
  case OP_JLT_REAL: {
    COMPARE_AND_JUMP(i32, <);
  }
  case OP_JLE_REAL: {
    COMPARE_AND_JUMP(i32, <=);
  }
    case OP_INT_TO_STR: {
    DECODE_A(instruction)
    
    u32 i = MAX_LEN_INT32;
    i32 v = (i32)locals[src1];
    char buffer[MAX_LEN_INT32];
    i32 n = v;
    bool neg = n < 0;
    USED(src2);

    if (neg)
      n = -n;

    do {
      buffer[--i] = radix_set[n % 10];
      n /= 10;
    } while (n > 0);
    if (neg)
      buffer[--i] = '-';
    /* Ensure at least one digit is written for 0 */
    if (v == 0)
      buffer[--i] = '0';

    /* Copy from buffer[i] to buffer + MAX_LEN_INT32 */
    locals[dest] = str_alloc(buffer + i, MAX_LEN_INT32 - i);

    return pc;
  }
  case OP_NAT_TO_STR: {
    DECODE_A(instruction)
    
    u32 v = (i32)locals[src1];
    char buffer[MAX_LEN_INT32];
    u32 n = v;
    u32 i = MAX_LEN_INT32;
    USED(src2);
    do {
      buffer[--i] = radix_set[n % 10];
      n /= 10;
    } while (n > 0);
    /* Ensure at least one digit is written for 0 */
    if (v == 0)
      buffer[--i] = '0';
    /* Copy from buffer[i] to buffer + MAX_LEN_INT32 */
    locals[dest] = str_alloc(buffer + i, MAX_LEN_INT32 - i);

    return pc;
  }
  case OP_REAL_TO_STR: {
    DECODE_A(instruction)
    
    u32 i = 0, j = 0;
    i32 q = (i32)locals[src1];
    char buffer[MAX_LEN_INT32];
    u32 int_part, frac_part;

    if (q < 0) {
      buffer[i++] = '-';
      q = -q;
    }

    int_part = q >> 16;
    frac_part = q & 0xFFFF;
    USED(src2);

    if (int_part == 0) {
      buffer[i++] = radix_set[0];
    } else {
      char tmp[16];
      i32 tmp_i = 0;
      while (int_part > 0) {
        tmp[tmp_i++] = radix_set[int_part % 10];
        int_part /= 10;
      }
      while (tmp_i > 0) {
        buffer[i++] = tmp[--tmp_i];
      }
    }

    buffer[i++] = '.';
    for (j = 0; j < 6; j++) {
      frac_part *= 10;
      buffer[i++] = radix_set[frac_part >> 16];
      frac_part &= 0xFFFF;           
    }

    locals[dest] = str_alloc(buffer + i, MAX_LEN_INT32 - i);
    
    return pc;
  }
  }

  /* something went very wrong */
  status = 255;
  return false;
}