bANS with support for variable L

#define L 48  // between 16 and 48
// The I/O is still 16b-based.

size_t bANSEncode(const uint8_t* in, size_t in_size,
                  uint8_t* const buf_start, uint8_t* const buf_end,
                  uint16_t p0) {
  const uint16_t q0 = 0x10000u - p0;
  const uint64_t threshold0 = (uint64_t)p0 << L;
  const uint64_t threshold1 = (uint64_t)q0 << L;
  uint16_t* buf = (uint16_t*)buf_end;
  uint64_t x = 1ull << L;
  int i;
  for (i = in_size - 1; i >= 0; --i) {
    if (x >= (in[i] ? threshold1 : threshold0)) {
      *--buf = x & 0xffff;
      x >>= 16;
    }
    if (in[i]) {
      x = ((x / q0) << 16) + (x % q0) + p0;
    } else {
      x = ((x / p0) << 16) + (x % p0) +  0;
    }
  }
  *--buf = (x >>  0) & 0xffff;
  *--buf = (x >> 16) & 0xffff;
  *--buf = (x >> 32) & 0xffff;
  *--buf = (x >> 48) & 0xffff;
  return buf_end - (uint8_t*)buf;
}

void bANSDecode(const uint16_t* ptr, uint8_t* out, size_t in_size, uint16_t p0) {
  int i;
  uint64_t x = 0;
  for (i = 0; i < 4; ++i) x = (x << 16) | *ptr++;
  for (i = 0; i < in_size; ++i) {
    if (x < (1ull << L)) {
      x = (x << 16) | *ptr++;   // decode forward
    }
    const uint16_t xfrac = x & 0xffff;
    const uint64_t x_new = p0 * (x >> 16);
    out[i] = (xfrac >= p0);
    if (xfrac < p0) {
      x = x_new + xfrac;
    } else {
      x -= x_new + p0;
    }
  }
}