diff --git a/include/benchmark/benchmark_api.h b/include/benchmark/benchmark_api.h
index 1e853e2c..47553221 100644
--- a/include/benchmark/benchmark_api.h
+++ b/include/benchmark/benchmark_api.h
@@ -231,7 +231,13 @@ BENCHMARK_UNUSED static int stream_init_anchor = InitializeStreams();
 #ifndef BENCHMARK_HAS_NO_INLINE_ASSEMBLY
 template <class Tp>
 inline BENCHMARK_ALWAYS_INLINE void DoNotOptimize(Tp const& value) {
+  // Clang doesn't like the 'X' constraint on `value` and certain GCC versions
+  // don't like the 'g' constraint. Attempt to placate them both.
+#if defined(__clang__)
   asm volatile("" : : "g"(value) : "memory");
+#else
+  asm volatile("" : : "i,r,m"(value) : "memory");
+#endif
 }
 // Force the compiler to flush pending writes to global memory. Acts as an
 // effective read/write barrier
diff --git a/test/CMakeLists.txt b/test/CMakeLists.txt
index d89135aa..b55612b4 100644
--- a/test/CMakeLists.txt
+++ b/test/CMakeLists.txt
@@ -1,6 +1,7 @@
 # Enable the tests
 
 find_package(Threads REQUIRED)
+include(CheckCXXCompilerFlag)
 
 # NOTE: Some tests use `<cassert>` to perform the test. Therefore we must
 # strip -DNDEBUG from the default CMake flags in DEBUG mode.
@@ -75,6 +76,11 @@ compile_benchmark_test(skip_with_error_test)
 add_test(skip_with_error_test skip_with_error_test --benchmark_min_time=0.01)
 
 compile_benchmark_test(donotoptimize_test)
+# Some of the issues with DoNotOptimize only occur when optimization is enabled
+check_cxx_compiler_flag(-O3 BENCHMARK_HAS_O3_FLAG)
+if (BENCHMARK_HAS_O3_FLAG)
+  set_target_properties(donotoptimize_test PROPERTIES COMPILE_FLAGS "-O3")
+endif()
 add_test(donotoptimize_test donotoptimize_test --benchmark_min_time=0.01)
 
 compile_benchmark_test(fixture_test)
diff --git a/test/donotoptimize_test.cc b/test/donotoptimize_test.cc
index b21187aa..a705654a 100644
--- a/test/donotoptimize_test.cc
+++ b/test/donotoptimize_test.cc
@@ -9,6 +9,22 @@ std::uint64_t double_up(const std::uint64_t x) __attribute__((const));
 std::uint64_t double_up(const std::uint64_t x) { return x * 2; }
 }
 
+// Using DoNotOptimize on types like BitRef seem to cause a lot of problems
+// with the inline assembly on both GCC and Clang.
+struct BitRef {
+  int index;
+  unsigned char &byte;
+
+public:
+  static BitRef Make() {
+    static unsigned char arr[2] = {};
+    BitRef b(1, arr[0]);
+    return b;
+  }
+private:
+  BitRef(int i, unsigned char& b) : index(i), byte(b) {}
+};
+
 int main(int, char*[]) {
   // this test verifies compilation of DoNotOptimize() for some types
 
@@ -29,5 +45,8 @@ int main(int, char*[]) {
 
   benchmark::DoNotOptimize(double_up(x));
 
-  return 0;
+  // These tests are to e
+  benchmark::DoNotOptimize(BitRef::Make());
+  BitRef lval = BitRef::Make();
+  benchmark::DoNotOptimize(lval);
 }