diff --git a/blake2-simd.cpp b/blake2-simd.cpp
index bd6ffb84..dc3f407a 100644
--- a/blake2-simd.cpp
+++ b/blake2-simd.cpp
@@ -11,6 +11,10 @@
 #include "misc.h"
 #include "blake2.h"
 
+#if (CRYPTOPP_SSE2_AVAILABLE)
+# include "emmintrin.h"
+#endif
+
 #if (CRYPTOPP_SSE42_AVAILABLE)
 # include "nmmintrin.h"
 #endif
diff --git a/gcm.cpp b/gcm.cpp
index e75c43a4..5341f36e 100644
--- a/gcm.cpp
+++ b/gcm.cpp
@@ -28,7 +28,7 @@
 
 NAMESPACE_BEGIN(CryptoPP)
 
-#if (CRYPTOPP_CLMUL_AVAILABLE)
+#if (CRYPTOPP_SSE2_AVAILABLE)
 # include "emmintrin.h"
 #endif
 
diff --git a/rijndael-simd.cpp b/rijndael-simd.cpp
index b14b3d5b..83c5dbf9 100644
--- a/rijndael-simd.cpp
+++ b/rijndael-simd.cpp
@@ -31,12 +31,10 @@
 #endif
 
 #if (CRYPTOPP_SSE41_AVAILABLE)
-// Hack... Clang conflates SSE4.1 and SSE4.2. Without __SSE4_2__,
-//   early Apple and LLVM compilers fail with "SSE4.2 instruction
-//   set not enabled" when "nmmintrin.h" is included. Its non-trivial
-//   for us to automatically add -msse4.2 for Apple Clang. We also want
-//   to avoid problems on low-end Atoms which have AES but lack SSE4.2.
-# if defined(__clang__)
+// Hack... GCC 4.8, LLVM Clang 3.5 and Apple Clang 6.0 conflates SSE4.1
+//   and SSE4.2. Without __SSE4_2__, early compilers fail with "SSE4.2
+//   instruction set not enabled" when "nmmintrin.h" is included.
+# if defined(__clang__) || defined(__GNUC__)
 #  define __SSE4_2__ 1
 # endif
 # include "nmmintrin.h"