Little better [#CLICKHOUSE-2].

dbms/src/Columns/ColumnsCommon.cpp

@@ -38,6 +38,8 @@ size_t countBytesInFilter(const IColumn::Filter & filt)
             | (static_cast<UInt64>(_mm_movemask_epi8(_mm_cmpgt_epi8(
                 _mm_loadu_si128(reinterpret_cast<const __m128i *>(pos + 48)),
                 zero16))) << 48));
+
+    /// TODO Add duff device for tail?
 #endif
 
     for (; pos < end; ++pos)
@@ -47,6 +49,44 @@ size_t countBytesInFilter(const IColumn::Filter & filt)
 }
 
 
+/** clang 4 generates better code than gcc 6.
+  * And both gcc and clang could not vectorize trivial loop by bytes automatically.
+  */
+bool memoryIsZero(const void * data, size_t size)
+{
+    const Int8 * pos = reinterpret_cast<const Int8 *>(data);
+    const Int8 * end = pos + size;
+
+#if __SSE2__
+    const __m128 zero16 = _mm_setzero_ps();
+    const Int8 * end64 = pos + size / 64 * 64;
+
+    for (; pos < end64; pos += 64)
+        if (_mm_movemask_ps(_mm_cmpneq_ps(
+                _mm_loadu_ps(reinterpret_cast<const float *>(pos)),
+                zero16))
+            | _mm_movemask_ps(_mm_cmpneq_ps(
+                _mm_loadu_ps(reinterpret_cast<const float *>(pos + 16)),
+                zero16))
+            | _mm_movemask_ps(_mm_cmpneq_ps(
+                _mm_loadu_ps(reinterpret_cast<const float *>(pos + 32)),
+                zero16))
+            | _mm_movemask_ps(_mm_cmpneq_ps(
+                _mm_loadu_ps(reinterpret_cast<const float *>(pos + 48)),
+                zero16)))
+            return false;
+
+    /// TODO Add duff device for tail?
+#endif
+
+    for (; pos < end; ++pos)
+        if (*pos)
+            return false;
+
+    return true;
+}
+
+
 namespace ErrorCodes
 {
     extern const int SIZES_OF_COLUMNS_DOESNT_MATCH;

dbms/src/Columns/ColumnsCommon.h

@@ -11,6 +11,9 @@ namespace DB
 /// Counts how many bytes of `filt` are greater than zero.
 size_t countBytesInFilter(const IColumn::Filter & filt);
 
+/// Returns true, if the memory contains only zeros.
+bool memoryIsZero(const void * data, size_t size);
+
 
 /// The general implementation of `filter` function for ColumnArray and ColumnString.
 template <typename T>

dbms/src/DataStreams/NullableAdapterBlockInputStream.cpp

@@ -1,8 +1,10 @@
 #include <DataStreams/NullableAdapterBlockInputStream.h>
 #include <Columns/ColumnNullable.h>
+#include <Columns/ColumnsCommon.h>
 #include <DataTypes/DataTypeNullable.h>
 #include <DataStreams/isConvertableTypes.h>
 
+
 namespace DB
 {
 
@@ -51,7 +53,7 @@ Block NullableAdapterBlockInputStream::readImpl()
                 const auto & nullable_type = static_cast<const DataTypeNullable &>(*elem.type);
 
                 const auto & null_map = nullable_col.getNullMap();
-                bool has_nulls = std::any_of(null_map.begin(), null_map.end(), [](UInt8 val){ return val == 1; });
+                bool has_nulls = !memoryIsZero(null_map.data(), null_map.size());
 
                 if (has_nulls)
                     throw Exception{"Cannot insert NULL value into non-nullable column",