diff --git a/CMakeLists.txt b/CMakeLists.txt
index 559291e8c34..06496bde59b 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -46,10 +46,6 @@ if (CMAKE_SYSTEM_PROCESSOR MATCHES "^(aarch64.*|AARCH64.*)")
 	set (AARCH64 1)
 endif ()
 
-if (NOT AARCH64)
-	set (MACHINE_FLAGS "-msse4 -mpopcnt")
-endif ()
-
 set (COMMON_WARNING_FLAGS "-Wall")	# -Werror is also added inside directories with our own code.
 set (CXX_WARNING_FLAGS "-Wnon-virtual-dtor")
 
@@ -110,14 +106,21 @@ if (PIPE)
 	set (COMPILER_FLAGS                   "${COMPILER_FLAGS} -pipe")
 endif ()
 
+include (cmake/test_cpu.cmake)
+
+option (ARCHNATIVE "Enable -march=native compiler flag" OFF)
+if (ARCHNATIVE)
+	set (COMPILER_FLAGS                   "${COMPILER_FLAGS} -march=native")
+endif ()
+
 
 set (CMAKE_BUILD_COLOR_MAKEFILE          ON)
-set (CMAKE_CXX_FLAGS                     "${CMAKE_CXX_FLAGS} ${COMPILER_FLAGS} -std=gnu++1y ${PLATFORM_EXTRA_CXX_FLAG} -fno-omit-frame-pointer ${COMMON_WARNING_FLAGS} ${CXX_WARNING_FLAGS} ${MACHINE_FLAGS} ${GLIBC_COMPATIBILITY_COMPILE_FLAGS}")
+set (CMAKE_CXX_FLAGS                     "${CMAKE_CXX_FLAGS} ${COMPILER_FLAGS} -std=gnu++1y ${PLATFORM_EXTRA_CXX_FLAG} -fno-omit-frame-pointer ${COMMON_WARNING_FLAGS} ${CXX_WARNING_FLAGS} ${GLIBC_COMPATIBILITY_COMPILE_FLAGS}")
 #set (CMAKE_CXX_FLAGS_RELEASE             "${CMAKE_CXX_FLAGS_RELEASE}")
 set (CMAKE_CXX_FLAGS_RELWITHDEBINFO      "${CMAKE_CXX_FLAGS_RELWITHDEBINFO} -O3")
 set (CMAKE_CXX_FLAGS_DEBUG               "${CMAKE_CXX_FLAGS_DEBUG} -O0 -g3 -ggdb3 -fno-inline")
 
-set (CMAKE_C_FLAGS                       "${CMAKE_C_FLAGS} ${COMPILER_FLAGS} -fno-omit-frame-pointer ${COMMON_WARNING_FLAGS} ${MACHINE_FLAGS} ${GLIBC_COMPATIBILITY_COMPILE_FLAGS}")
+set (CMAKE_C_FLAGS                       "${CMAKE_C_FLAGS} ${COMPILER_FLAGS} -fno-omit-frame-pointer ${COMMON_WARNING_FLAGS} ${GLIBC_COMPATIBILITY_COMPILE_FLAGS}")
 #set (CMAKE_C_FLAGS_RELEASE               "${CMAKE_C_FLAGS_RELEASE}")
 set (CMAKE_C_FLAGS_RELWITHDEBINFO        "${CMAKE_C_FLAGS_RELWITHDEBINFO} -O3")
 set (CMAKE_C_FLAGS_DEBUG                 "${CMAKE_C_FLAGS_DEBUG} -O0 -g3 -ggdb3 -fno-inline")
diff --git a/cmake/find_readline_edit.cmake b/cmake/find_readline_edit.cmake
index 3bf70c6a4a0..a3f0d4ddd71 100644
--- a/cmake/find_readline_edit.cmake
+++ b/cmake/find_readline_edit.cmake
@@ -62,7 +62,3 @@ check_cxx_source_runs("
 		append_history(1,nullptr);
 	}
 " HAVE_READLINE_HISTORY)
-
-#if (HAVE_READLINE_HISTORY)
-#	add_definitions (-D HAVE_READLINE_HISTORY)
-#endif ()
diff --git a/cmake/test_cpu.cmake b/cmake/test_cpu.cmake
new file mode 100644
index 00000000000..699d8fe7b92
--- /dev/null
+++ b/cmake/test_cpu.cmake
@@ -0,0 +1,63 @@
+# https://software.intel.com/sites/landingpage/IntrinsicsGuide/
+
+include (CheckCXXSourceCompiles)
+
+# gcc -dM -E -mno-sse2 - < /dev/null | sort > gcc-dump-nosse2
+# gcc -dM -E -msse2 - < /dev/null | sort > gcc-dump-sse2
+#define __SSE2__ 1
+#define __SSE2_MATH__ 1
+
+# gcc -dM -E -msse4.1 - < /dev/null | sort > gcc-dump-sse41
+#define __SSE4_1__ 1
+
+set (TEST_FLAG "-msse4.1")
+set (CMAKE_REQUIRED_FLAGS "${TEST_FLAG}")
+check_cxx_source_compiles("
+	#include <smmintrin.h>
+	int main() {
+		_mm_insert_epi8(__m128i(), 0, 0);
+		return 0;
+	}
+" HAVE_SSE41)
+if (HAVE_SSE41)
+	set (COMPILER_FLAGS "${COMPILER_FLAGS} ${TEST_FLAG}")
+endif ()
+
+
+# gcc -dM -E -msse4.2 - < /dev/null | sort > gcc-dump-sse42
+#define __SSE4_2__ 1
+
+set (TEST_FLAG "-msse4.2")
+set (CMAKE_REQUIRED_FLAGS "${TEST_FLAG}")
+check_cxx_source_compiles("
+	#include <nmmintrin.h>
+	int main() {
+		_mm_crc32_u64(0, 0);
+		return 0;
+	}
+" HAVE_SSE42)
+if (HAVE_SSE42)
+	set (COMPILER_FLAGS "${COMPILER_FLAGS} ${TEST_FLAG}")
+endif ()
+
+
+# gcc -dM -E -mpopcnt - < /dev/null | sort > gcc-dump-popcnt
+#define __POPCNT__ 1
+
+set (TEST_FLAG "-mpopcnt")
+
+set (CMAKE_REQUIRED_FLAGS "${TEST_FLAG}")
+check_cxx_source_compiles("
+	int main() {
+		__builtin_popcountll(0);
+		return 0;
+	}
+" HAVE_POPCNT)
+
+if (HAVE_POPCNT AND NOT AARCH64)
+	set (COMPILER_FLAGS "${COMPILER_FLAGS} ${TEST_FLAG}")
+endif ()
+
+set (CMAKE_REQUIRED_FLAGS "")
+
+# TODO: add here sse3 test if you want use it
diff --git a/dbms/include/DB/Columns/ColumnVector.h b/dbms/include/DB/Columns/ColumnVector.h
index 6359b253525..8fbf920f02a 100644
--- a/dbms/include/DB/Columns/ColumnVector.h
+++ b/dbms/include/DB/Columns/ColumnVector.h
@@ -12,7 +12,7 @@
 
 #include <DB/Columns/IColumn.h>
 
-#if defined(__x86_64__)
+#if __SSE2__
 	#include <emmintrin.h>
 #endif
 
@@ -357,7 +357,7 @@ public:
 		const UInt8 * filt_end = filt_pos + size;
 		const T * data_pos = &data[0];
 
-#if defined(__x86_64__)
+#if __SSE2__
 		/** Чуть более оптимизированная версия.
 		 * Исходит из допущения, что часто куски последовательно идущих значений
 		 *  полностью проходят или полностью не проходят фильтр.
diff --git a/dbms/include/DB/Common/BlockFilterCreator.h b/dbms/include/DB/Common/BlockFilterCreator.h
index aa17ec04953..75a6c736cf9 100644
--- a/dbms/include/DB/Common/BlockFilterCreator.h
+++ b/dbms/include/DB/Common/BlockFilterCreator.h
@@ -5,7 +5,7 @@
 
 #include <type_traits>
 
-#if defined(__x86_64__)
+#if __SSE2__
 	#define LIBDIVIDE_USE_SSE2 1
 #endif
 
diff --git a/dbms/include/DB/Common/StringSearcher.h b/dbms/include/DB/Common/StringSearcher.h
index 5540aeb3abe..94b1933897d 100644
--- a/dbms/include/DB/Common/StringSearcher.h
+++ b/dbms/include/DB/Common/StringSearcher.h
@@ -7,7 +7,7 @@
 #include <stdint.h>
 #include <string.h>
 
-#if defined(__x86_64__)
+#if __SSE4_1__
 	#include <smmintrin.h>
 #endif
 
@@ -60,7 +60,7 @@ private:
 	UInt8 l{};
 	UInt8 u{};
 
-#if defined(__x86_64__)
+#if __SSE4_1__
 	/// vectors filled with `l` and `u`, for determining leftmost position of the first symbol
 	__m128i patl, patu;
 	/// lower and uppercase vectors of first 16 characters of `needle`
@@ -99,7 +99,7 @@ public:
 			u = u_seq[0];
 		}
 
-#if defined(__x86_64__)
+#if __SSE4_1__
 		/// for detecting leftmost position of the first symbol
 		patl = _mm_set1_epi8(l);
 		patu = _mm_set1_epi8(u);
@@ -160,7 +160,7 @@ public:
 	{
 		static const Poco::UTF8Encoding utf8;
 
-#if defined(__x86_64__)
+#if __SSE4_1__
 		if (page_safe(pos))
 		{
 			const auto v_haystack = _mm_loadu_si128(reinterpret_cast<const __m128i *>(pos));
@@ -225,7 +225,7 @@ public:
 
 		while (haystack < haystack_end)
 		{
-#if defined(__x86_64__)
+#if __SSE4_1__
 			if (haystack + n <= haystack_end && page_safe(haystack))
 			{
 				const auto v_haystack = _mm_loadu_si128(reinterpret_cast<const __m128i *>(haystack));
@@ -330,7 +330,7 @@ private:
 	UInt8 l{};
 	UInt8 u{};
 
-#if defined(__x86_64__)
+#if __SSE4_1__
 	/// vectors filled with `l` and `u`, for determining leftmost position of the first symbol
 	__m128i patl, patu;
 	/// lower and uppercase vectors of first 16 characters of `needle`
@@ -348,7 +348,7 @@ public:
 		l = static_cast<UInt8>(std::tolower(*needle));
 		u = static_cast<UInt8>(std::toupper(*needle));
 
-#if defined(__x86_64__)
+#if __SSE4_1__
 		patl = _mm_set1_epi8(l);
 		patu = _mm_set1_epi8(u);
 
@@ -372,7 +372,7 @@ public:
 
 	bool compare(const UInt8 * pos) const
 	{
-#if defined(__x86_64__)
+#if __SSE4_1__
 		if (page_safe(pos))
 		{
 			const auto v_haystack = _mm_loadu_si128(reinterpret_cast<const __m128i *>(pos));
@@ -424,7 +424,7 @@ public:
 
 		while (haystack < haystack_end)
 		{
-#if defined(__x86_64__)
+#if __SSE4_1__
 			if (haystack + n <= haystack_end && page_safe(haystack))
 			{
 				const auto v_haystack = _mm_loadu_si128(reinterpret_cast<const __m128i *>(haystack));
@@ -516,7 +516,7 @@ private:
 	/// first character in `needle`
 	UInt8 first{};
 
-#if defined(__x86_64__)
+#if __SSE4_1__
 	/// vector filled `first` for determining leftmost position of the first symbol
 	__m128i pattern;
 	/// vector of first 16 characters of `needle`
@@ -533,7 +533,7 @@ public:
 
 		first = *needle;
 
-#if defined(__x86_64__)
+#if __SSE4_1__
 		pattern = _mm_set1_epi8(first);
 
 		auto needle_pos = needle;
@@ -554,7 +554,7 @@ public:
 
 	bool compare(const UInt8 * pos) const
 	{
-#if defined(__x86_64__)
+#if __SSE4_1__
 		if (page_safe(pos))
 		{
 			const auto v_haystack = _mm_loadu_si128(reinterpret_cast<const __m128i *>(pos));
@@ -604,7 +604,7 @@ public:
 
 		while (haystack < haystack_end)
 		{
-#if defined(__x86_64__)
+#if __SSE4_1__
 			if (haystack + n <= haystack_end && page_safe(haystack))
 			{
 				/// find first character
diff --git a/dbms/include/DB/Common/memcpySmall.h b/dbms/include/DB/Common/memcpySmall.h
index 5828c046a40..b35a6343294 100644
--- a/dbms/include/DB/Common/memcpySmall.h
+++ b/dbms/include/DB/Common/memcpySmall.h
@@ -3,9 +3,7 @@
 #include <string.h>
 #include <DB/Core/Defines.h>
 
-
-#if defined(__x86_64__)
-
+#if __SSE2__
 #include <emmintrin.h>
 
 
diff --git a/dbms/include/DB/Core/StringRef.h b/dbms/include/DB/Core/StringRef.h
index 6b8e34b2a29..93f2184f384 100644
--- a/dbms/include/DB/Core/StringRef.h
+++ b/dbms/include/DB/Core/StringRef.h
@@ -5,7 +5,7 @@
 #include <functional>
 #include <ostream>
 
-#if defined(__x86_64__)
+#if __SSE2__
 	#include <emmintrin.h>
 #endif
 
@@ -35,7 +35,7 @@ using StringRefs = std::vector<StringRef>;
 
 using UInt64 = DB::UInt64;
 
-#if defined(__x86_64__)
+#if __SSE2__
 
 /** Сравнение строк на равенство.
   * Подход является спорным и выигрывает не во всех случаях.
@@ -128,7 +128,7 @@ inline bool operator== (StringRef lhs, StringRef rhs)
 	if (lhs.size == 0)
 		return true;
 
-#if defined(__x86_64__)
+#if __SSE2__
 	return memequalSSE2Wide(lhs.data, rhs.data, lhs.size);
 #else
 	return 0 == memcmp(lhs.data, rhs.data, lhs.size);
@@ -169,7 +169,7 @@ struct StringRefHash64
 	}
 };
 
-#if defined(__x86_64__)
+#if __SSE4_2__
 
 #ifdef __SSE4_1__
 #include <smmintrin.h>
diff --git a/dbms/include/DB/Functions/FunctionsArithmetic.h b/dbms/include/DB/Functions/FunctionsArithmetic.h
index e613001b7a1..d4eca307799 100644
--- a/dbms/include/DB/Functions/FunctionsArithmetic.h
+++ b/dbms/include/DB/Functions/FunctionsArithmetic.h
@@ -943,7 +943,7 @@ template <> struct FunctionUnaryArithmeticMonotonicity<NameBitNot>
 
 /// Оптимизации для целочисленного деления на константу.
 
-#if defined(__x86_64__)
+#if __SSE2__
 	#define LIBDIVIDE_USE_SSE2 1
 #endif
 
@@ -981,7 +981,7 @@ struct DivideIntegralByConstantImpl
 		const A * a_end = a_pos + size;
 		ResultType * c_pos = &c[0];
 
-#if defined(__x86_64__)
+#if __SSE2__
 		static constexpr size_t values_per_sse_register = 16 / sizeof(A);
 		const A * a_end_sse = a_pos + size / values_per_sse_register * values_per_sse_register;
 
diff --git a/dbms/include/DB/Functions/FunctionsRound.h b/dbms/include/DB/Functions/FunctionsRound.h
index ed808f47155..1b62ea5a6d6 100644
--- a/dbms/include/DB/Functions/FunctionsRound.h
+++ b/dbms/include/DB/Functions/FunctionsRound.h
@@ -5,7 +5,7 @@
 #include <type_traits>
 #include <array>
 
-#if defined(__x86_64__)
+#if __SSE4_1__
 	#include <smmintrin.h>
 #endif
 
@@ -268,7 +268,7 @@ namespace DB
 		}
 	};
 
-#if defined(__x86_64__)
+#if __SSE4_1__
 	template <typename T>
 	class BaseFloatRoundingComputation;
 
diff --git a/dbms/include/DB/Functions/FunctionsString.h b/dbms/include/DB/Functions/FunctionsString.h
index 3429bf09cf6..a2d7741a896 100644
--- a/dbms/include/DB/Functions/FunctionsString.h
+++ b/dbms/include/DB/Functions/FunctionsString.h
@@ -16,9 +16,8 @@
 #include <DB/Functions/IFunction.h>
 #include <ext/range.hpp>
 
-#if defined(__x86_64__)
+#if __SSE2__
 	#include <emmintrin.h>
-	#include <nmmintrin.h>
 #endif
 
 
@@ -232,7 +231,7 @@ private:
 	{
 		const auto flip_case_mask = 'A' ^ 'a';
 
-#if defined(__x86_64__)
+#if __SSE2__
 		const auto bytes_sse = sizeof(__m128i);
 		const auto src_end_sse = src_end - (src_end - src) % bytes_sse;
 
@@ -393,7 +392,7 @@ private:
 
 	static void array(const UInt8 * src, const UInt8 * src_end, UInt8 * dst)
 	{
-#if defined(__x86_64__)
+#if __SSE2__
 		const auto bytes_sse = sizeof(__m128i);
 		auto src_end_sse = src + (src_end - src) / bytes_sse * bytes_sse;
 
diff --git a/dbms/src/Columns/ColumnsCommon.cpp b/dbms/src/Columns/ColumnsCommon.cpp
index 2fcd7ddeb17..1acf6c5c8e8 100644
--- a/dbms/src/Columns/ColumnsCommon.cpp
+++ b/dbms/src/Columns/ColumnsCommon.cpp
@@ -1,4 +1,4 @@
-#if defined(__x86_64__)
+#if __SSE2__
 	#include <emmintrin.h>
 #endif
 
@@ -20,7 +20,7 @@ size_t countBytesInFilter(const IColumn::Filter & filt)
 	const Int8 * pos = reinterpret_cast<const Int8 *>(&filt[0]);
 	const Int8 * end = pos + filt.size();
 
-#if defined(__x86_64__)
+#if __SSE2__ && __POPCNT__
 	const __m128i zero16 = _mm_setzero_si128();
 	const Int8 * end64 = pos + filt.size() / 64 * 64;
 
@@ -95,7 +95,7 @@ void filterArraysImpl(
 		memcpy(&res_elems[elems_size_old], &src_elems[offset], size * sizeof(T));
 	};
 
-#if defined(__x86_64__)
+#if __SSE2__
 	const __m128i zero_vec = _mm_setzero_si128();
 	static constexpr size_t SIMD_BYTES = 16;
 	const auto filt_end_aligned = filt_pos + size / SIMD_BYTES * SIMD_BYTES;
diff --git a/dbms/src/DataTypes/DataTypeString.cpp b/dbms/src/DataTypes/DataTypeString.cpp
index 9b6f166425a..80395786552 100644
--- a/dbms/src/DataTypes/DataTypeString.cpp
+++ b/dbms/src/DataTypes/DataTypeString.cpp
@@ -10,7 +10,7 @@
 #include <DB/IO/WriteHelpers.h>
 #include <DB/IO/VarInt.h>
 
-#if defined(__x86_64__)
+#if __SSE2__
 	#include <emmintrin.h>
 #endif
 
@@ -125,7 +125,7 @@ static NO_INLINE void deserializeBinarySSE2(ColumnString::Chars_t & data, Column
 
 		if (size)
 		{
-#if defined(__x86_64__)
+#if __SSE2__
 			/// Оптимистичная ветка, в которой возможно более эффективное копирование.
 			if (offset + 16 * UNROLL_TIMES <= data.allocated_size() && istr.position() + size + 16 * UNROLL_TIMES <= istr.buffer().end())
 			{
diff --git a/dbms/src/IO/ReadHelpers.cpp b/dbms/src/IO/ReadHelpers.cpp
index dbb107ce09f..0d9c609957a 100644
--- a/dbms/src/IO/ReadHelpers.cpp
+++ b/dbms/src/IO/ReadHelpers.cpp
@@ -1,7 +1,3 @@
-#if defined(__x86_64__)
-	#include <emmintrin.h>
-#endif
-
 #include <sstream>
 
 #include <mysqlxx/Manip.h>
diff --git a/dbms/src/IO/WriteBufferValidUTF8.cpp b/dbms/src/IO/WriteBufferValidUTF8.cpp
index ed2fdebb500..e2b7c4b1bb4 100644
--- a/dbms/src/IO/WriteBufferValidUTF8.cpp
+++ b/dbms/src/IO/WriteBufferValidUTF8.cpp
@@ -2,8 +2,8 @@
 #include <DB/IO/WriteBufferValidUTF8.h>
 #include <DB/Core/Types.h>
 
-#ifdef __x86_64__
-#include <emmintrin.h>
+#if __SSE2__
+	#include <emmintrin.h>
 #endif
 
 
@@ -69,7 +69,7 @@ void WriteBufferValidUTF8::nextImpl()
 
 	while (p < pos)
 	{
-#ifdef __x86_64__
+#if __SSE2__
 		/// Fast skip of ASCII
 		static constexpr size_t SIMD_BYTES = 16;
 		const char * simd_end = p + (pos - p) / SIMD_BYTES * SIMD_BYTES;
diff --git a/dbms/src/IO/tests/mempbrk.cpp b/dbms/src/IO/tests/mempbrk.cpp
index 537f01858d7..41a6d99e83e 100644
--- a/dbms/src/IO/tests/mempbrk.cpp
+++ b/dbms/src/IO/tests/mempbrk.cpp
@@ -1,7 +1,3 @@
-#if defined(__x86_64__)
-	#include <nmmintrin.h>
-#endif
-
 #include <string>
 #include <iostream>
 #include <iomanip>
diff --git a/dbms/src/IO/tests/parse_int_perf2.cpp b/dbms/src/IO/tests/parse_int_perf2.cpp
index 427b261f03c..691e73eda05 100644
--- a/dbms/src/IO/tests/parse_int_perf2.cpp
+++ b/dbms/src/IO/tests/parse_int_perf2.cpp
@@ -1,7 +1,3 @@
-#if defined(__x86_64__)
-	#include <emmintrin.h>
-#endif
-
 #include <iostream>
 #include <iomanip>
 
@@ -11,22 +7,6 @@
 #include <DB/Common/Stopwatch.h>
 
 
-#if defined(__x86_64__)
-std::ostream & operator<< (std::ostream & ostr, const __m128i vec)
-{
-	char digits[16];
-	_mm_store_si128(reinterpret_cast<__m128i *>(digits), vec);
-
-	ostr << "{";
-	for (size_t i = 0; i < 16; ++i)
-		ostr << (i ? ", " : "") << static_cast<int>(digits[i]);
-	ostr << "}";
-	
-	return ostr;
-}
-#endif
-
-
 namespace test
 {
 	template <typename T>
diff --git a/dbms/src/Interpreters/tests/hash_map_string_2.cpp b/dbms/src/Interpreters/tests/hash_map_string_2.cpp
index a121d80cc01..f060867343d 100644
--- a/dbms/src/Interpreters/tests/hash_map_string_2.cpp
+++ b/dbms/src/Interpreters/tests/hash_map_string_2.cpp
@@ -15,7 +15,7 @@
 #include <DB/Common/HashTable/HashMap.h>
 #include <DB/Interpreters/AggregationCommon.h>
 
-#if defined(__x86_64__)
+#if __SSE4_1__
 	#include <smmintrin.h>
 #endif
 
@@ -77,7 +77,7 @@ DefineStringRef(StringRef_Compare16_1_byMemcmp)
 DefineStringRef(StringRef_Compare16_1_byUInt64_logicAnd)
 DefineStringRef(StringRef_Compare16_1_byUInt64_bitAnd)
 
-#if defined(__x86_64__)
+#if __SSE4_1__
 DefineStringRef(StringRef_Compare16_1_byIntSSE)
 DefineStringRef(StringRef_Compare16_1_byFloatSSE)
 DefineStringRef(StringRef_Compare16_1_bySSE4)
@@ -196,7 +196,7 @@ inline bool compare_byUInt64_bitAnd(const char * p1, const char * p2)
 		 & (reinterpret_cast<const UInt64 *>(p1)[1] == reinterpret_cast<const UInt64 *>(p2)[1]);
 }
 
-#if defined(__x86_64__)
+#if __SSE4_1__
 
 inline bool compare_byIntSSE(const char * p1, const char * p2)
 {
@@ -263,7 +263,7 @@ inline bool memequal(const char * p1, const char * p2, size_t size)
 }
 
 
-#if defined(__x86_64__)
+#if __SSE4_1__
 
 inline bool memequal_sse41(const char * p1, const char * p2, size_t size)
 {
@@ -514,7 +514,7 @@ Op(byMemcmp)
 Op(byUInt64_logicAnd)
 Op(byUInt64_bitAnd)
 
-#if defined(__x86_64__)
+#if __SSE4_1__
 
 Op(byIntSSE)
 Op(byFloatSSE)
@@ -642,7 +642,7 @@ int main(int argc, char ** argv)
 	if (!m || m == 5) bench<StringRef_Compare16_1_byMemcmp>			(data, "StringRef_Compare16_1_byMemcmp");
 	if (!m || m == 6) bench<StringRef_Compare16_1_byUInt64_logicAnd>(data, "StringRef_Compare16_1_byUInt64_logicAnd");
 	if (!m || m == 7) bench<StringRef_Compare16_1_byUInt64_bitAnd>	(data, "StringRef_Compare16_1_byUInt64_bitAnd");
-#if defined(__x86_64__)
+#if __SSE4_1__
 	if (!m || m == 8) bench<StringRef_Compare16_1_byIntSSE>			(data, "StringRef_Compare16_1_byIntSSE");
 	if (!m || m == 9) bench<StringRef_Compare16_1_byFloatSSE>		(data, "StringRef_Compare16_1_byFloatSSE");
 	if (!m || m == 10) bench<StringRef_Compare16_1_bySSE4>			(data, "StringRef_Compare16_1_bySSE4");
diff --git a/dbms/src/Interpreters/tests/hash_map_string_3.cpp b/dbms/src/Interpreters/tests/hash_map_string_3.cpp
index f05a427642c..8712b8817c2 100644
--- a/dbms/src/Interpreters/tests/hash_map_string_3.cpp
+++ b/dbms/src/Interpreters/tests/hash_map_string_3.cpp
@@ -18,7 +18,7 @@
 #include <DB/Common/HashTable/HashMap.h>
 #include <DB/Interpreters/AggregationCommon.h>
 
-#if defined(__x86_64__)
+#if __SSE4_1__
 	#include <smmintrin.h>
 #endif
 
@@ -139,7 +139,7 @@ struct FastHash64
 };
 
 
-#if defined(__x86_64__)
+#if __SSE4_1__
 
 struct CrapWow
 {
@@ -229,7 +229,7 @@ struct SimpleHash
 
 		if (size < 8)
 		{
-#if defined(__x86_64__)
+#if __SSE4_1__
 			return hashLessThan8(x.data, x.size);
 #endif
 		}
@@ -266,7 +266,7 @@ struct VerySimpleHash
 
 		if (size < 8)
 		{
-#if defined(__x86_64__)
+#if __SSE4_1__
 			return hashLessThan8(x.data, x.size);
 #endif
 		}
@@ -316,7 +316,7 @@ struct MetroHash64
 };
 
 
-#if defined(__x86_64__)
+#if __SSE4_1__
 
 /*struct CRC32Hash
 {
@@ -466,7 +466,7 @@ int main(int argc, char ** argv)
 	if (!m || m == 2) bench<StringRef_CompareMemcmp, FastHash64>	(data, "StringRef_FastHash64");
 	if (!m || m == 3) bench<StringRef_CompareMemcmp, SimpleHash>	(data, "StringRef_SimpleHash");
 
-#if defined(__x86_64__)
+#if __SSE4_1__
 	if (!m || m == 4) bench<StringRef_CompareMemcmp, CrapWow>		(data, "StringRef_CrapWow");
 	if (!m || m == 5) bench<StringRef_CompareMemcmp, CRC32Hash>		(data, "StringRef_CRC32Hash");
 	if (!m || m == 6) bench<StringRef_CompareMemcmp, CRC32ILPHash>	(data, "StringRef_CRC32ILPHash");
diff --git a/dbms/src/Interpreters/tests/hash_map_string_small.cpp b/dbms/src/Interpreters/tests/hash_map_string_small.cpp
index 534dd25102b..e4fd8661085 100644
--- a/dbms/src/Interpreters/tests/hash_map_string_small.cpp
+++ b/dbms/src/Interpreters/tests/hash_map_string_small.cpp
@@ -64,7 +64,7 @@ inline bool operator==(SmallStringRef lhs, SmallStringRef rhs)
 	if (lhs.size == 0)
 		return true;
 
-#if __x86_64__
+#if __SSE2__
 	return memequalSSE2Wide(lhs.data(), rhs.data(), lhs.size);
 #else
 	return false;
diff --git a/libs/libcommon/include/common/find_first_symbols.h b/libs/libcommon/include/common/find_first_symbols.h
index 52ae6951ea6..33c9a96cf77 100644
--- a/libs/libcommon/include/common/find_first_symbols.h
+++ b/libs/libcommon/include/common/find_first_symbols.h
@@ -1,6 +1,9 @@
 #pragma once
 
-#if defined(__x86_64__)
+#if __SSE2__
+	#include <emmintrin.h>
+#endif
+#if __SSE4_2__
 	#include <nmmintrin.h>
 #endif
 
@@ -35,7 +38,7 @@ inline bool is_in(char x)
 	return x == s0 || is_in<s1, tail...>(x);
 }
 
-#if defined(__x86_64__)
+#if __SSE2__
 template <char s0>
 inline __m128i mm_is_in(__m128i bytes)
 {
@@ -56,7 +59,7 @@ inline __m128i mm_is_in(__m128i bytes)
 template <char... symbols>
 inline const char * find_first_symbols_sse2(const char * begin, const char * end)
 {
-#if defined(__x86_64__)
+#if __SSE2__
 	for (; begin + 15 < end; begin += 16)
 	{
 		__m128i bytes = _mm_loadu_si128(reinterpret_cast<const __m128i *>(begin));
@@ -83,7 +86,7 @@ template <size_t num_chars,
 	char c13 = 0, char c14 = 0, char c15 = 0, char c16 = 0>
 inline const char * find_first_symbols_sse42_impl(const char * begin, const char * end)
 {
-#if defined(__x86_64__)
+#if __SSE4_2__
 #define MODE (_SIDD_UBYTE_OPS | _SIDD_CMP_EQUAL_ANY | _SIDD_LEAST_SIGNIFICANT)
 	__m128i set = _mm_setr_epi8(c01, c02, c03, c04, c05, c06, c07, c08, c09, c10, c11, c12, c13, c14, c15, c16);
 
@@ -131,8 +134,10 @@ inline const char * find_first_symbols_sse42(const char * begin, const char * en
 template <char... symbols>
 inline const char * find_first_symbols(const char * begin, const char * end)
 {
+#if __SSE4_2__
 	if (sizeof...(symbols) >= 5)
 		return detail::find_first_symbols_sse42<symbols...>(begin, end);
 	else
+#endif
 		return detail::find_first_symbols_sse2<symbols...>(begin, end);
 }
diff --git a/utils/fill-factor/main.cpp b/utils/fill-factor/main.cpp
index f5da10bbb96..a924848a035 100644
--- a/utils/fill-factor/main.cpp
+++ b/utils/fill-factor/main.cpp
@@ -1,7 +1,7 @@
 #include <iostream>
 #include <iomanip>
 
-#if defined(__x86_64__)
+#if __SSE2__
 #include <emmintrin.h>
 #endif
 
@@ -15,7 +15,7 @@
 
 int main(int argc, char ** argv)
 {
-#if defined(__x86_64__)
+#if __SSE2__
 	try
 	{
 		DB::ReadBufferFromFileDescriptor in(STDIN_FILENO);