#include "vm.h"

#define FRAME_HEADER_SIZE 12

u32 pc;       /* program counter */
u32 cp;       /* code pointer */
u32 mp;       /* memory pointer */
u32 fp;       /* frame pointer */
u32 flag;     /* flag */
u8 interrupt; /* device interrupt */
u32 *code;    /* code */
u8 *mem;      /* memory */

bool step_vm() {

  u32 instruction = code[pc++];
  u8 opcode = DECODE_OP(instruction);

  switch (opcode) {
  case OP_HALT: {
    /* no need to decode, all are zeros */
    return false;
  }
  case OP_CALL: {
    DECODE_A(instruction)
    u32 fn_ptr, local_count, return_ptr;

    u32 rd = fp + dest;
    u32 r1 = fp + src1;
    u32 r2 = fp + src2;

    fn_ptr = READ_U32(rd);
    local_count = READ_U32(r1);
    return_ptr = READ_U32(r2);

    /* push parents frame value to reset the heap to */
    WRITE_U32(mp, fp);
    mp += 4;
    /* push return address to child frame */
    WRITE_U32(mp, pc + 1);
    mp += 4;
    /* push local address to return the value to */
    WRITE_U32(mp, return_ptr);
    mp += 4;
    /* now set the frame pointer, where the locals start */
    fp = mp;
    /* move mp by count many locals */
    mp += (4 * local_count);
    /* jump to dest_ptr */
    pc = fn_ptr;
    return true;
  }
  case OP_RETURN: {
    DECODE_B(instruction)
    u32 i, size = 0;
    u32 return_local = fp + dest;
    u32 return_value = READ_U32(return_local);
    bool is_ptr = (((u16)(1)) << 15) & imm;
    bool replaces_value = (((u16)(1)) << 14) & imm;

    /* reset mp to saved mp, use header size to get "real" start of frame */
    u32 frame_start = fp - FRAME_HEADER_SIZE;
    u32 parent_fp = READ_U32(frame_start);
    u32 return_address = READ_U32(frame_start + 4);
    u32 parent_local_return_address = READ_U32(frame_start + 8);

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

    /* copy value to end of mp if it is a pointer */
    if (is_ptr) {
      WRITE_U32(parent_local_return_address, mp);
      size = READ_U32(return_value);
      WRITE_U32(mp, size);
      mp += 4;
      for (i = 0; i < size; i++) {
        u8 value = READ_U8(return_value + i);
        WRITE_U8(mp, value);
        mp++;
      }
    } else {
    /* otherwise just write the return value to its location */
      WRITE_U32(parent_local_return_address, return_value);
    }

    /* jump to parent frame */
    pc = return_address;
    return true;
  }
  case OP_SYSCALL: {
    DECODE_A(instruction)
    u32 id = fp + dest;
    u32 args = fp + src1;
    u32 mem_ptr = fp + src2;
    flag = syscall(id, args, mem_ptr);
    return true;
  }
  case OP_LOAD_IMM: {
    DECODE_B(instruction)
    u32 rd = fp + dest;
    WRITE_U32(rd, imm);
    return true;
  }
  case OP_LOAD_UPPER_IMM: {
    DECODE_B(instruction)
    u32 rd = fp + dest;
    u32 value = READ_U32(rd);
    WRITE_U32(rd, (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, ldest;
    u32 rd = fp + dest;
    u32 r1 = fp + src1;
    USED(src2);
    ldest = READ_U32(rd);
    WRITE_U32(ldest, mp);
    size = READ_U32(r1);
    WRITE_U32(mp, size);
    mp += (size + 4);
    return true;
  }
  case OP_MEM_CPY_8: {
    DECODE_A(instruction)
    u32 i, count, mdest, msrc;

    u32 rd = fp + dest;
    u32 r1 = fp + src1;
    u32 r2 = fp + src2;

    mdest = READ_U32(rd);
    msrc = READ_U32(r1);
    count = READ_U32(r2);

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

    for (i = 0; i < count; i++) {
      u8 value = READ_U8(mdest + i);
      WRITE_U8(msrc + i, value);
    }

    flag = 0;
    return true;
  }
  case OP_MEM_CPY_16: {
    DECODE_A(instruction)
    u32 i, count, mdest, msrc;

    u32 rd = fp + dest;
    u32 r1 = fp + src1;
    u32 r2 = fp + src2;

    mdest = READ_U32(rd);
    msrc = READ_U32(r1);
    count = READ_U32(r2);

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

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

    flag = 0;
    return true;
  }
  case OP_MEM_CPY_32: {
    DECODE_A(instruction)
    u32 i, count, mdest, msrc;

    u32 rd = fp + dest;
    u32 r1 = fp + src1;
    u32 r2 = fp + src2;

    mdest = READ_U32(rd);
    msrc = READ_U32(r1);
    count = READ_U32(r2);

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

    for (i = 0; i < count; i++) {
      u32 value = READ_U32(mdest + i);
      WRITE_U32(msrc + i, value);
    }

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

    u32 rd = fp + dest;
    u32 r1 = fp + src1;
    u32 r2 = fp + src2;

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

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

    start = rd;
    end = start + count;

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

    for (i = start; i < end; i++) {
      WRITE_U8(i, value);
    }

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

    u32 rd = fp + dest;
    u32 r1 = fp + src1;
    u32 r2 = fp + src2;

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

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

    start = rd;
    end = start + count;

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

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

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

    u32 rd = fp + dest;
    u32 r1 = fp + src1;
    u32 r2 = fp + src2;

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

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

    start = rd;
    end = start + count;

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

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

    flag = 0;
    return true;
  }
  case OP_MOV: {
    DECODE_A(instruction)
    u32 rd = fp + dest;
    u32 r1 = fp + src1;
    u32 value = READ_U32(r1);
    USED(src2);
    WRITE_U32(rd, value);
    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)
    u32 rd = fp + dest;
    u32 r1 = fp + src1;
    u32 r2 = fp + src2;

    i32 src1_whole = (i32)READ_U32(r1) >> 16;
    i32 src2_whole = (i32)READ_U32(r2) >> 16;

    i32 src1_decimal = (i32)READ_U32(r1) & 16;
    i32 src2_decimal = (i32)READ_U32(r2) & 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;
    WRITE_U32(rd, result);
    return true;
  }
  case OP_DIV_REAL: {
    DECODE_A(instruction)
    i32 result;
    u32 rd = fp + dest;
    u32 r1 = fp + src1;
    u32 r2 = fp + src2;

    i32 src1_val = (i32)READ_U32(r1);
    i32 src2_val = (i32)READ_U32(r2);

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

    result = src1_val * src2_reciprocal;
    result <<= 1;
    WRITE_U32(rd, result);
    return true;
  }
  case OP_INT_TO_REAL: {
    DECODE_A(instruction)
    u32 rd = fp + dest;
    u32 r1 = fp + src1;
    i32 result = (i32)READ_U32(r1) << 16;
    USED(src2);
    WRITE_U32(rd, result);
    return true;
  }
  case OP_INT_TO_NAT: {
    DECODE_A(instruction)
    u32 rd = fp + dest;
    u32 r1 = fp + src1;
    u32 result = (u32)READ_U32(r1);
    USED(src2);
    WRITE_U32(rd, result);
    return true;
  }
  case OP_NAT_TO_REAL: {
    DECODE_A(instruction)
    u32 rd = fp + dest;
    u32 r1 = fp + src1;
    i32 result = ((i32)READ_U32(r1) << 16);
    USED(src2);
    WRITE_U32(rd, result);
    return true;
  }
  case OP_NAT_TO_INT: {
    DECODE_A(instruction)
    u32 rd = fp + dest;
    u32 r1 = fp + src1;
    i32 result = ((i32)READ_U32(r1));
    USED(src2);
    WRITE_U32(rd, result);
    return true;
  }
  case OP_REAL_TO_INT: {
    DECODE_A(instruction)
    u32 rd = fp + dest;
    u32 r1 = fp + src1;
    i32 result = ((i32)READ_U32(r1) >> 16);
    USED(src2);
    WRITE_U32(rd, result);
    return true;
  }
  case OP_REAL_TO_NAT: {
    DECODE_A(instruction)
    u32 rd = fp + dest;
    u32 r1 = fp + src1;
    u32 result = ((u32)READ_U32(r1) >> 16);
    USED(src2);
    WRITE_U32(rd, 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: {
    DECODE_C(instruction)
    pc = imm;
    return true;
  }
  case OP_JMP_ABS: {
    DECODE_A(instruction)
    u32 rd = fp + dest;
    u32 jmp_dest = READ_U32(rd);
    if (jmp_dest > cp) {
      flag = 1;
      return true;
    }
    USED(src1);
    USED(src2);

    pc = jmp_dest;
    return true;
  }
  case OP_JMP_OFF: {
    DECODE_A(instruction)
    u32 rd = fp + dest;
    u32 r1 = fp + src1;

    u32 jmp_dest = READ_U32(rd) + READ_U32(r1);
    if (jmp_dest > cp) {
      flag = 1;
      return true;
    }
    USED(src2);

    pc = jmp_dest;
    return true;
  }
  case OP_JMP_FLAG: {
    DECODE_A(instruction)
    u32 mask;
    u32 rd = fp + dest;
    u32 jmp_dest = READ_U32(rd);
    if (jmp_dest > cp) {
      flag = 1;
      return true;
    }
    USED(src1);
    USED(src2);

    mask = -(u32)(flag == 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, <=);
  }
  }

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