/*********************************************************************
 * Filename:      bitvec.h
 *
 * Description:   Bit vectors implemented as uint64_t arrays, size
 *                fixed at compile time (in weylconfig.h).  Supports
 *                efficient set operations: Union, difference, count.
 *                Uses SSE4 64-bit popcount instruction.
 *
 * Author:        David Dumas <david@dumas.io>
 *
 * This program is free software distributed under the MIT license.
 * See the file LICENSE for details.
 ********************************************************************/


#ifndef __BITVEC_H__
#define __BITVEC_H__ 1

// #define _GNU_SOURCE /* enable ffsll */
// #include <string.h>

#include <inttypes.h>

#include <stdio.h>
#include <stdlib.h>

/* For _mm_popcnt_u64 */
#include <x86intrin.h>

// FIRSTBITS(n) only yields useful result when 0 <= n < 64
#define FIRSTBITS(n) ((1l << (n)) - 1l)
#define ALLBITS ((uint64_t)-1)

typedef struct {
  uint64_t v[BV_QWORD_RANK];
} bitvec_t;

/*
static inline unsigned long bv_hash(bitvec_t S)
{
  unsigned long h = 5381;
  int i,j;
  uint8_t x;

  for (i = 0; i < BV_QWORD_RANK*4; i++) {
    x = (S.v[i/8] >> (8*(i % 8))) & 0xff;
    h = ((h << 5) + h) ^ x;
  }
  return h;
}

static inline int bv_equal(bitvec_t x,bitvec_t y)
{
  int i;
  for (i=0; i < BV_QWORD_RANK; i++) {
    if (x.v[i] != y.v[i]) {
      return 0;
    }
  }
  return 1;
  //  return ((x.v[0] == y.v[0]) && (x.v[1] == y.v[1]));
}

static inline void bv_difference(bitvec_t *pile, bitvec_t toremove)
{
  int i;
  for (i=0; i < BV_QWORD_RANK; i++) {
    pile->v[i] &= ~toremove.v[i];
  }
}
*/

/* /\* static inline *\/ int old_popcount64(uint64_t x) */
/* { */
/*   int i; */
/*   for (i = 0; x; x >>= 1) { */
/*     i += x & 1; */
/*   } */
/*   return i; */
/* } */

/*
static inline int popcount64(uint64_t x)
{
  return _mm_popcnt_u64(x);
}

static inline int bv_popcount(bitvec_t x)
{
  int i;
  int accum=0;
  for (i=0; i<BV_QWORD_RANK; i++) {
    accum += popcount64(x.v[i]);
  }
  return accum;
}
*/

static inline void bv_clear_bit(bitvec_t *x, int k)
{
  x->v[k/64] &= ~((uint64_t)1 << (k % 64));
}

static inline void bv_set_bit(bitvec_t *x, int k)
{
  x->v[k/64] |= ((uint64_t)1 << (k % 64));
}

static inline int bv_get_bit(const bitvec_t *x, int k)
{
  return (x->v[k/64] >> (k % 64)) & 0x1;
}

static inline void bv_clear(bitvec_t *x)
{
  int i;
  for (i=0;i<BV_QWORD_RANK;i++)
    x->v[i] = 0;
}

static inline int bv_is_zero(const bitvec_t *x)
{
  int i;
  for (i=0;i<BV_QWORD_RANK;i++)
    if (x->v[i])
      return 0;

  return 1;
}

static inline void bv_print(FILE *f, const bitvec_t *x, int len)
{
  for(int i = 0; i < len; i++) {
    fputc(bv_get_bit(x, i) ? '1' : '0', f);
    if(i % 64 == 63)
      fputc('-',f);
  }
}

static inline void bv_union(const bitvec_t *x, const bitvec_t *y, bitvec_t *result)
{
  int i;
  for (i=0; i < BV_QWORD_RANK; i++) {
    result->v[i] = x->v[i] | y->v[i];
  }
}

static inline void bv_intersection(const bitvec_t *x, const bitvec_t *y, bitvec_t *result)
{
  int i;
  for (i=0; i < BV_QWORD_RANK; i++) {
    result->v[i] = x->v[i] & y->v[i];
  }
}

static inline int bv_disjoint(const bitvec_t *x, const bitvec_t *y)
{
  for(int i = 0; i < BV_QWORD_RANK; i++)
    if(x->v[i] & y->v[i])
      return 0;

  return 1;
}

static inline int bv_full(const bitvec_t *x, int len)
{
  int i;
  for(i = 0; i < len/64; i++)
    if(x->v[i] != ALLBITS)
      return 0;

  return (x->v[i] & FIRSTBITS(len%64)) == FIRSTBITS(len%64);
}

// set bits in range start...end (including start and excluding end)
static inline void bv_set_range(bitvec_t *x, int start, int end)
{
  if(start/64 == end/64)
    x->v[start/64] |= ~FIRSTBITS(start % 64) & FIRSTBITS(end % 64);
  else {
    x->v[start/64] |= ~FIRSTBITS(start % 64);
    for(int i = start/64 + 1; i < end/64; i++)
      x->v[i] = ALLBITS;
    x->v[end/64] |= FIRSTBITS(end % 64);
  }
}

// find least significant 0 bit starting from position start (included)
static inline int bv_next_zero(const bitvec_t *x, int start)
{
  int position;

  position = ffsll(~(x->v[start/64] | FIRSTBITS(start % 64)));

  if(position)
    return (start/64)*64 + position - 1; // found zero in same chunk

  for(int i = start/64 + 1; i < BV_QWORD_RANK; i++) {
    position = ffsll(~x->v[i]);
    if(position) // found a 0
      return i*64 + position - 1;
  }

  return BV_QWORD_RANK; // found nothing
}

static inline void bv_copy(const bitvec_t *from, bitvec_t *to)
{
  for(int i = 0; i < BV_QWORD_RANK; i++)
    to->v[i] = from->v[i];
}

#endif /* __BITVEC_H__ */