From af7700e124db7b9aeb7c80d32d258adaad44875e Mon Sep 17 00:00:00 2001
From: Kim Walisch <kim.walisch@gmail.com>
Date: Wed, 19 Jun 2024 11:19:26 +0200
Subject: [PATCH] Tune AVX512 code

---
 include/Sieve.hpp | 14 +++++++++-----
 1 file changed, 9 insertions(+), 5 deletions(-)

diff --git a/include/Sieve.hpp b/include/Sieve.hpp
index 2fc9486d..df52396a 100644
--- a/include/Sieve.hpp
+++ b/include/Sieve.hpp
@@ -208,14 +208,18 @@ class Sieve
       // Compute this for loop using AVX512.
       // for (i = start_idx + 1; i < stop_idx; i++)
       //   cnt += popcnt64(sieve64[i]);
-      do {
-        __mmask8 mask = (i + 8 < stop_idx) ? 0xff : (__mmask8) _bzhi_u64(0xff, stop_idx - i);
-        __m512i vec = _mm512_maskz_loadu_epi64(mask , &sieve64[i]);
+      //
+      for (; i + 8 < stop_idx; i += 8)
+      {
+        __m512i vec = _mm512_loadu_epi64(&sieve64[i]);
         vec = _mm512_popcnt_epi64(vec);
         vcnt = _mm512_add_epi64(vcnt, vec);
-        i += 8;
       }
-      while (i < stop_idx);
+
+      __mmask8 mask = (__mmask8) _bzhi_u64(0xff, stop_idx - i);
+      __m512i 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);