diff --git a/GNUmakefile b/GNUmakefile
index 89950acf..a14b3daf 100755
--- a/GNUmakefile
+++ b/GNUmakefile
@@ -238,9 +238,12 @@ ifeq ($(findstring -DCRYPTOPP_DISABLE_SSSE3,$(CXXFLAGS)),)
     SSSE3_FLAG = -mssse3
   endif
 ifeq ($(findstring -DCRYPTOPP_DISABLE_SSE4,$(CXXFLAGS)),)
+  HAVE_SSE4 = $(shell echo | $(CXX) -x c++ $(CXXFLAGS) -msse4.1 -dM -E - 2>/dev/null | $(GREP) -i -c __SSE4_1__)
+  ifeq ($(HAVE_SSE4),1)
+    BLAKE2_FLAG = -msse4.1
+  endif
   HAVE_SSE4 = $(shell echo | $(CXX) -x c++ $(CXXFLAGS) -msse4.2 -dM -E - 2>/dev/null | $(GREP) -i -c __SSE4_2__)
   ifeq ($(HAVE_SSE4),1)
-    BLAKE2_FLAG = -msse4.2
     CRC_FLAG = -msse4.2
   endif
 ifeq ($(findstring -DCRYPTOPP_DISABLE_AESNI,$(CXXFLAGS)),)
diff --git a/blake2-simd.cpp b/blake2-simd.cpp
index a3f05c76..6ad53f97 100644
--- a/blake2-simd.cpp
+++ b/blake2-simd.cpp
@@ -13,16 +13,17 @@
 
 // Uncomment for benchmarking C++ against SSE2 or NEON.
 // Do so in both blake2.cpp and blake2-simd.cpp.
-// #undef CRYPTOPP_SSE42_AVAILABLE
+// #undef CRYPTOPP_SSE41_AVAILABLE
 // #undef CRYPTOPP_ARM_NEON_AVAILABLE
 
 #if !(defined(__ARM_NEON) || defined(_MSC_VER))
 # undef CRYPTOPP_ARM_NEON_AVAILABLE
 #endif
 
-#if (CRYPTOPP_SSE42_AVAILABLE)
+#if (CRYPTOPP_SSE41_AVAILABLE)
 # include <emmintrin.h>
-# include <nmmintrin.h>
+# include <tmmintrin.h>
+# include <smmintrin.h>
 #endif
 
 #if (CRYPTOPP_ARM_NEON_AVAILABLE)
@@ -75,7 +76,7 @@ const word64 BLAKE2B_IV[8] = {
 
 ANONYMOUS_NAMESPACE_END
 
-#if CRYPTOPP_SSE42_AVAILABLE
+#if CRYPTOPP_SSE41_AVAILABLE
 void BLAKE2_Compress32_SSE4(const byte* input, BLAKE2_State<word32, false>& state)
 {
   __m128i row1, row2, row3, row4;
@@ -1605,7 +1606,7 @@ void BLAKE2_Compress64_SSE4(const byte* input, BLAKE2_State<word64, true>& state
   _mm_storeu_si128(M128_CAST(&state.h[4]), _mm_xor_si128(_mm_loadu_si128(CONST_M128_CAST(&state.h[4])), row2l));
   _mm_storeu_si128(M128_CAST(&state.h[6]), _mm_xor_si128(_mm_loadu_si128(CONST_M128_CAST(&state.h[6])), row2h));
 }
-#endif  // CRYPTOPP_SSE42_AVAILABLE
+#endif  // CRYPTOPP_SSE41_AVAILABLE
 
 // Disable NEON for Cortex-A53 and A57. Also see http://github.com/weidai11/cryptopp/issues/367
 #if CRYPTOPP_BOOL_ARM32 && CRYPTOPP_ARM_NEON_AVAILABLE
diff --git a/blake2.cpp b/blake2.cpp
index c4be4621..580d941e 100644
--- a/blake2.cpp
+++ b/blake2.cpp
@@ -14,19 +14,13 @@ NAMESPACE_BEGIN(CryptoPP)
 
 // Uncomment for benchmarking C++ against SSE2 or NEON.
 // Do so in both blake2.cpp and blake2-simd.cpp.
-// #undef CRYPTOPP_SSE42_AVAILABLE
+// #undef CRYPTOPP_SSE41_AVAILABLE
 // #undef CRYPTOPP_ARM_NEON_AVAILABLE
 
-// Apple Clang 6.0/Clang 3.5 does not have SSSE3 intrinsics
-//   http://llvm.org/bugs/show_bug.cgi?id=20213
-#if (defined(CRYPTOPP_APPLE_CLANG_VERSION) && (CRYPTOPP_APPLE_CLANG_VERSION <= 60000)) || (defined(CRYPTOPP_LLVM_CLANG_VERSION) && (CRYPTOPP_LLVM_CLANG_VERSION <= 30500))
-# undef CRYPTOPP_SSE42_AVAILABLE
-#endif
-
 void BLAKE2_Compress32_CXX(const byte* input, BLAKE2_State<word32, false>& state);
 void BLAKE2_Compress64_CXX(const byte* input, BLAKE2_State<word64, true>& state);
 
-#if CRYPTOPP_SSE42_AVAILABLE
+#if CRYPTOPP_SSE41_AVAILABLE
 extern void BLAKE2_Compress32_SSE4(const byte* input, BLAKE2_State<word32, false>& state);
 extern void BLAKE2_Compress64_SSE4(const byte* input, BLAKE2_State<word64, true>& state);
 #endif
@@ -95,8 +89,8 @@ typedef void (*pfnCompress64)(const byte*, BLAKE2_State<word64, true>&);
 
 pfnCompress64 InitializeCompress64Fn()
 {
-#if CRYPTOPP_SSE42_AVAILABLE
-    if (HasSSE42())
+#if CRYPTOPP_SSE41_AVAILABLE
+    if (HasSSE41())
         return &BLAKE2_Compress64_SSE4;
     else
 #endif
@@ -110,8 +104,8 @@ pfnCompress64 InitializeCompress64Fn()
 
 pfnCompress32 InitializeCompress32Fn()
 {
-#if CRYPTOPP_SSE42_AVAILABLE
-    if (HasSSE42())
+#if CRYPTOPP_SSE41_AVAILABLE
+    if (HasSSE41())
         return &BLAKE2_Compress32_SSE4;
     else
 #endif