const modulo optimization

dbms/src/Functions/modulo.cpp

@@ -61,8 +61,23 @@ struct ModuloByConstantImpl
 
         /// Here we failed to make the SSE variant from libdivide give an advantage.
         size_t size = a.size();
-        for (size_t i = 0; i < size; ++i)
-            c[i] = a[i] - (a[i] / divider) * b; /// NOTE: perhaps, the division semantics with the remainder of negative numbers is not preserved.
+
+        /// strict aliasing optimization for char like arrays
+        auto * __restrict src = a.data();
+        auto * __restrict dst = c.data();
+
+        if (b & (b - 1))
+        {
+            for (size_t i = 0; i < size; ++i)
+                dst[i] = src[i] - (src[i] / divider) * b; /// NOTE: perhaps, the division semantics with the remainder of negative numbers is not preserved.
+        }
+        else
+        {
+            // gcc libdivide doesn't work well for pow2 division
+            auto mask = b - 1;
+            for (size_t i = 0; i < size; ++i)
+                dst[i] = src[i] & mask;
+        }
     }
 };
 

dbms/tests/performance/modulo.xml

@@ -0,0 +1,17 @@
+<test>
+    <type>loop</type>
+
+    <stop_conditions>
+        <any_of>
+            <iterations>10</iterations>
+        </any_of>
+    </stop_conditions>
+
+    <main_metric>
+        <min_time />
+    </main_metric>
+
+    <query>SELECT number % 128 FROM numbers(300000000) FORMAT Null</query>
+    <query>SELECT number % 255 FROM numbers(300000000) FORMAT Null</query>
+    <query>SELECT number % 256 FROM numbers(300000000) FORMAT Null</query>
+</test>