var x [128]uint32 // give x its own cache line
	sink = &x
	b.RunParallel(func(pb *testing.PB) {
		i := uint32(0)
		for pb.Next() {
			atomic.Or32(&x[63], i)
			i++
		}
	})
}

func BenchmarkOr64(b *testing.B) {
	var x [128]uint64 // give x its own cache line
	sink = &x
	for i := 0; i < b.N; i++ {
		atomic.Or64(&x[63], uint64(i))
	}
}

func BenchmarkOr64Parallel(b *testing.B) {