Improve AVX512 code

include/Sieve.hpp

@@ -203,28 +203,27 @@ class Sieve
     else
     {
       uint64_t i = start_idx + 1;
-      uint64_t cnt = popcnt64(sieve64[start_idx] & m1);
-      __m512i vcnt = _mm512_setzero_si512();
+      uint64_t start_bits = sieve64[start_idx] & m1;
+      uint64_t stop_bits = sieve64[stop_idx] & m2;
+      __m512i vec = _mm512_set_epi64(0, 0, 0, 0, 0, 0, stop_bits, start_bits);
+      __m512i vcnt = _mm512_popcnt_epi64(vec);
 
       // Compute this for loop using AVX512.
       // for (i = start_idx + 1; i < stop_idx; i++)
       //   cnt += popcnt64(sieve64[i]);
       //
       for (; i + 8 < stop_idx; i += 8)
       {
-        __m512i vec = _mm512_loadu_epi64(&sieve64[i]);
+        vec = _mm512_loadu_epi64(&sieve64[i]);
         vec = _mm512_popcnt_epi64(vec);
         vcnt = _mm512_add_epi64(vcnt, vec);
       }
 
       __mmask8 mask = (__mmask8) _bzhi_u64(0xff, stop_idx - i);
-      __m512i vec = _mm512_maskz_loadu_epi64(mask, &sieve64[i]);
+      vec = _mm512_maskz_loadu_epi64(mask, &sieve64[i]);
       vec = _mm512_popcnt_epi64(vec);
       vcnt = _mm512_add_epi64(vcnt, vec);
-
-      cnt += _mm512_reduce_add_epi64(vcnt);
-      cnt += popcnt64(sieve64[stop_idx] & m2);
-      return cnt;
+      return _mm512_reduce_add_epi64(vcnt);
     }
   }