diff --git a/CMakeLists.txt b/CMakeLists.txt
index 625a434..e552f7b 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -142,6 +142,19 @@ if (HAVE_PTHREAD_CANCEL)
 	add_definitions(/DHAVE_PTHREAD_CANCEL)
 endif()
 
+include(CheckCXXSourceCompiles)
+
+check_cxx_source_compiles("int main(){ return __builtin_clzll(1);}" HAVE_BUILTIN_CLZLL)
+check_cxx_source_compiles("#include <intrin.h>\n#pragma intrinsic(_BitScanReverse64)\nint main(){unsigned long r;_BitScanReverse64(&r,1);return r;}" HAVE_BITSCANREVERSE64)
+
+if (HAVE_BUILTIN_CLZLL)
+	add_definitions(/DHAVE_BUILTIN_CLZLL)
+endif()
+
+if (HAVE_BITSCANREVERSE64)
+	add_definitions(/DHAVE_BITSCANREVERSE64)
+endif()
+
 add_executable(${CMAKE_PROJECT_NAME} ${HEADERS} ${SOURCES})
 
 if (STATIC_BINARY)
diff --git a/src/stratum_server.cpp b/src/stratum_server.cpp
index 5d18564..d14a190 100644
--- a/src/stratum_server.cpp
+++ b/src/stratum_server.cpp
@@ -562,19 +562,18 @@ enum HashMaxValue : uint64_t {
 	value = hash_uncompress(std::numeric_limits<uint16_t>::max())
 };
 
-static constexpr FORCEINLINE uint16_t hash_compress(uint64_t h)
+static FORCEINLINE uint16_t hash_compress(uint64_t h)
 {
-	if (h > HashMaxValue::value) {
-		h = HashMaxValue::value;
+	if (h <= HashValue::mask) {
+		return static_cast<uint16_t>(h);
 	}
 
-	uint64_t shift = 0;
-	while (h > HashValue::mask) {
-		h >>= 1;
-		shift += (1 << HashValue::bits);
+	if (h >= HashMaxValue::value) {
+		return std::numeric_limits<uint16_t>::max();
 	}
 
-	return static_cast<uint16_t>(shift | h);
+	const uint64_t shift = bsr(h) - (HashValue::bits - 1);
+	return static_cast<uint16_t>((shift << HashValue::bits) | (h >> shift));
 }
 
 }
diff --git a/src/util.cpp b/src/util.cpp
index 7b261a3..b96fe8d 100644
--- a/src/util.cpp
+++ b/src/util.cpp
@@ -384,4 +384,43 @@ bool resolve_host(std::string& host, bool& is_v6)
 
 RandomDeviceSeed RandomDeviceSeed::instance;
 
+struct BSR8
+{
+	uint8_t data[256];
+
+	static constexpr BSR8 init() {
+		BSR8 result = { 55 };
+
+		for (int i = 1; i < 256; ++i) {
+			int x = i;
+			result.data[i] = 63;
+			while (x < 0x80) {
+				--result.data[i];
+				x <<= 1;
+			}
+		}
+
+		return result;
+	}
+};
+
+static constexpr BSR8 bsr8_table = BSR8::init();
+
+NOINLINE uint64_t bsr_reference(uint64_t x)
+{
+	uint32_t y = static_cast<uint32_t>(x);
+
+	uint64_t n0 = (x == y) ? 0 : 32;
+	y = static_cast<uint32_t>(x >> n0);
+	n0 ^= 32;
+
+	const uint64_t n1 = (y & 0xFFFF0000UL) ? 0 : 16;
+	y <<= n1;
+
+	const uint64_t n2 = (y & 0xFF000000UL) ? 0 : 8;
+	y <<= n2;
+
+	return bsr8_table.data[y >> 24] - n0 - n1 - n2;
+}
+
 } // namespace p2pool
diff --git a/src/util.h b/src/util.h
index 453496f..7a95874 100644
--- a/src/util.h
+++ b/src/util.h
@@ -203,6 +203,22 @@ FORCEINLINE uint64_t seconds_since_epoch()
 	return duration_cast<seconds>(steady_clock::now().time_since_epoch()).count();
 }
 
+uint64_t bsr_reference(uint64_t x);
+
+#ifdef HAVE_BUILTIN_CLZLL
+#define bsr(x) (63 - __builtin_clzll(x))
+#elif defined HAVE_BITSCANREVERSE64
+#pragma intrinsic(_BitScanReverse64)
+FORCEINLINE uint64_t bsr(uint64_t x)
+{
+	unsigned long index;
+	_BitScanReverse64(&index, x);
+	return index;
+}
+#else
+#define bsr bsr_reference
+#endif
+
 } // namespace p2pool
 
 namespace robin_hood {
diff --git a/tests/CMakeLists.txt b/tests/CMakeLists.txt
index 9f3fc54..295097d 100644
--- a/tests/CMakeLists.txt
+++ b/tests/CMakeLists.txt
@@ -74,7 +74,7 @@ set(SOURCES
 	src/keccak_tests.cpp
 	src/main.cpp
 	src/pool_block_tests.cpp
-	src/varint_tests.cpp
+	src/util_tests.cpp
 	src/wallet_tests.cpp
 	../external/src/cryptonote/crypto-ops-data.c
 	../external/src/cryptonote/crypto-ops.c
@@ -160,6 +160,19 @@ endif()
 
 add_definitions(/DZMQ_STATIC /DP2POOL_LOG_DISABLE)
 
+include(CheckCXXSourceCompiles)
+
+check_cxx_source_compiles("int main(){ return __builtin_clzll(1);}" HAVE_BUILTIN_CLZLL)
+check_cxx_source_compiles("#include <intrin.h>\n#pragma intrinsic(_BitScanReverse64)\nint main(){unsigned long r;_BitScanReverse64(&r,1);return r;}" HAVE_BITSCANREVERSE64)
+
+if (HAVE_BUILTIN_CLZLL)
+	add_definitions(/DHAVE_BUILTIN_CLZLL)
+endif()
+
+if (HAVE_BITSCANREVERSE64)
+	add_definitions(/DHAVE_BITSCANREVERSE64)
+endif()
+
 add_executable(${CMAKE_PROJECT_NAME} ${HEADERS} ${SOURCES})
 target_link_libraries(${CMAKE_PROJECT_NAME} debug ${ZMQ_LIBRARY_DEBUG} debug ${UV_LIBRARY_DEBUG} optimized ${ZMQ_LIBRARY} optimized ${UV_LIBRARY} ${LIBS})
 add_custom_command(TARGET ${CMAKE_PROJECT_NAME} POST_BUILD COMMAND ${CMAKE_COMMAND} -E copy_if_different "${CMAKE_SOURCE_DIR}/src/crypto_tests.txt" $<TARGET_FILE_DIR:${CMAKE_PROJECT_NAME}>)
diff --git a/tests/src/varint_tests.cpp b/tests/src/util_tests.cpp
similarity index 94%
rename from tests/src/varint_tests.cpp
rename to tests/src/util_tests.cpp
index 8ed5815..9e572b3 100644
--- a/tests/src/varint_tests.cpp
+++ b/tests/src/util_tests.cpp
@@ -21,7 +21,7 @@
 
 namespace p2pool {
 
-TEST(varint, read_write)
+TEST(util, varint)
 {
 	std::vector<uint8_t> v;
 	v.reserve(16);
@@ -82,4 +82,12 @@ TEST(varint, read_write)
 	ASSERT_EQ(readVarint(buf2, buf2 + 1, check), nullptr);
 }
 
+TEST(util, bsr)
+{
+	for (uint64_t i = 0, x = 1; i <= 63; ++i, x <<= 1) {
+		ASSERT_EQ(bsr(x), i);
+		ASSERT_EQ(bsr_reference(x), i);
+	}
+}
+
 }