Add AVX512 trailing bytes optimization

kimwalisch · Jun 26, 2024 · adf413a · adf413a
1 parent 8fdabaa
commit adf413a
Showing 1 changed file with 7 additions and 0 deletions.
diff --git a/libpopcnt.h b/libpopcnt.h
@@ -553,8 +553,11 @@ static inline uint64_t popcnt(const void* data, uint64_t size)
       const uint64_t* ptr64 = (const uint64_t*)(ptr + i);
       cnt += popcnt_avx512(ptr64, (size - i) / 8);
       i = size - size % 8;
+
       if (i == size)
         return cnt;
+      else
+        goto trailing_bytes;
     }
 #endif
 
@@ -573,6 +576,10 @@ static inline uint64_t popcnt(const void* data, uint64_t size)
     }
 #endif
 
+#if defined(LIBPOPCNT_HAVE_AVX512)
+  trailing_bytes:
+#endif
+
 #if defined(LIBPOPCNT_HAVE_POPCNT)
   /* 
    * The user has compiled without -mpopcnt.