#include <string.h>

typedef unsigned vu __attribute__ ((vector_size (UNROLL*sizeof(unsigned))));

unsigned keylocks(unsigned keys[], unsigned nkeys, unsigned locks[], unsigned nlocks)
{
  unsigned i, j;
  unsigned part1 = 0;
  vu part0;
  int nlocks1=(nlocks+UNROLL-1)&(-UNROLL);
  int nkeys1;
  int locks1[nlocks1];
  int keys1[nkeys];
  for (i=0; i<nlocks; i++)
    locks1[i]=locks[i];
  for (; i<nlocks1; i++)
    locks1[i]=~0;
  /* special-casing for 0 keys, to avoid counting the extra locks */
  for (i=0, nkeys1=0; i<nkeys; i++)
    if (keys[i] != 0)
      keys1[nkeys1++]=keys[i];
    else
      part1+=nlocks;
  for (i=0; i<UNROLL; i++)
    part0[i]=0;
  for (i=0; i<nlocks1; i+=UNROLL) {
    vu lock = *(vu *)(locks1+i);
    for (j=0; j<nkeys1; j++) {
      part0 -= (lock & keys1[j])==0;
    }
  }
  for (i=0; i<UNROLL; i++)
    part1 += part0[i];
  return part1;
}