#include <iostream>
#include <cstdio>
#include <cstdlib>
#include <string.h>
using namespace std;

#include "MOA.h"
#include <emmintrin.h>

int main(int argc, char **argv)
{
  int print;
  int andor;
  int bytes, bufsize;
  int i, j;

  uint8_t *buf1, *buf2, *buf3;  
  __m128i *b1, *b2, *b3;

  /* usage: and_or bytes AND|OR|and|or|none seed print(Y|N) */

  if (argc != 5 || (strcmp(argv[2], "AND") != 0 && strcmp(argv[2], "OR") != 0  && 
                    strcmp(argv[2], "and") != 0 && strcmp(argv[2], "or") != 0  &&
                    strcmp(argv[2], "none") != 0)) {
    fprintf(stderr, "usage: and_or bytes AND|OR|and|or|none seed print(Y|N)\n");
    exit(1);
  }

  /* Set the variables from the command line */

  bytes = atoi(argv[1]);
  andor = argv[2][0];
  print = (argv[4][0] == 'Y');
  MOA_Seed(atoi(argv[3]));
  
  /* Pad the number of bytes so that it's a multiple of 16, and allocate three buffers: */

  bufsize = bytes;
  if (bufsize % 16 != 0) bufsize += (16 - bufsize%16);
  buf1 = (uint8_t *) malloc(bufsize);
  buf2 = (uint8_t *) malloc(bufsize);
  buf3 = (uint8_t *) malloc(bufsize);

  /* Use MOA.c to fill buf1 and buf2 with random bytes */

  MOA_Fill_Random_Region (buf1, bytes);
  MOA_Fill_Random_Region (buf2, bytes);

  /* This lets you access the three buffers as 128-bit vectors. */

  b1 = (__m128i *) buf1;
  b2 = (__m128i *) buf2;
  b3 = (__m128i *) buf3;

  /* Use SIMD to do AND. */

  if (andor == 'A') {
    for (i = 0; i < bytes; i += 16) {
      *b3 = _mm_and_si128(*b1, *b2);
      b1++;
      b2++;
      b3++;
    }

  /* Use SIMD to do OR. */

  } else if (andor == 'O') {
    for (i = 0; i < bytes; i += 16) {
      *b3 = _mm_or_si128(*b1, *b2);
      b1++;
      b2++;
      b3++;
    }

  /* Do the bit operations inefficiently on the bytes. */

  } else if (andor == 'a') {
    for (i = 0; i < bytes; i++) buf3[i] = (buf1[i] & buf2[i]);
  } else if (andor == 'o') {
    for (i = 0; i < bytes; i++) buf3[i] = (buf1[i] | buf2[i]);
  }

  /* Print everything out. */

  if (print) {
    printf("B1: ");
    for (i = 0; i < bytes; i++) {
      for (j = 0; j < 8; j++) printf("%d", (buf1[i] >> j) & 1);
      printf(" ");
    }
    printf("\n");
    printf("B2: ");
    for (i = 0; i < bytes; i++) {
      for (j = 0; j < 8; j++) printf("%d", (buf2[i] >> j) & 1);
      printf(" ");
    }
    printf("\n");
    printf("B3: ");
    for (i = 0; i < bytes; i++) {
      for (j = 0; j < 8; j++) printf("%d", (buf3[i] >> j) & 1);
      printf(" ");
    }
    printf("\n");
  }
  return 0;
}