Add CRYPTOPP_SLOW_ARMV8_SHIFT for Aarch32 and Aarch64

Both BLAKE2 and SPECK slow down when using NEON/ASIMD. When just BLAKE2 experienced the issue, it was a one-off problem. Its now wider than a one-off, so add the formal define

blake2-simd.cpp

@@ -16,6 +16,12 @@
 // #undef CRYPTOPP_SSE41_AVAILABLE
 // #undef CRYPTOPP_ARM_NEON_AVAILABLE
 
+// Disable NEON/ASIMD for Cortex-A53 and A57. The shifts are too slow and C/C++ is about
+// 3 cpb faster than NEON/ASIMD. Also see http://github.com/weidai11/cryptopp/issues/367.
+#if (defined(__aarch32__) || defined(__aarch64__)) && defined(CRYPTOPP_SLOW_ARMV8_SHIFT)
+# undef CRYPTOPP_ARM_NEON_AVAILABLE
+#endif
+
 #if !(defined(__ARM_NEON) || defined(_MSC_VER))
 # undef CRYPTOPP_ARM_NEON_AVAILABLE
 #endif
@@ -43,7 +49,7 @@ NAMESPACE_BEGIN(CryptoPP)
 inline __m128i MM_SET_EPI64X(const word64 a, const word64 b)
 {
     const word64 t[2] = {b,a}; __m128i r;
-    ::memcpy(&r, t, sizeof(t));
+    std::memcpy(&r, t, sizeof(t));
     return r;
 }
 #else
@@ -1600,8 +1606,7 @@ void BLAKE2_Compress64_SSE4(const byte* input, BLAKE2_State<word64, true>& state
 }
 #endif  // CRYPTOPP_SSE41_AVAILABLE
 
-// Disable NEON for Cortex-A53 and A57. Also see http://github.com/weidai11/cryptopp/issues/367
+#if CRYPTOPP_ARM_NEON_AVAILABLE
-#if CRYPTOPP_BOOL_ARM32 && CRYPTOPP_ARM_NEON_AVAILABLE
 void BLAKE2_Compress32_NEON(const byte* input, BLAKE2_State<word32, false>& state)
 {
     #define BLAKE2S_LOAD_MSG_0_1(buf) \
@@ -2179,6 +2184,6 @@ void BLAKE2_Compress64_NEON(const byte* input, BLAKE2_State<word64, true>& state
     vst1q_u64(&state.h[4], veorq_u64(h2, veorq_u64(row2l, row4l)));
     vst1q_u64(&state.h[6], veorq_u64(h3, veorq_u64(row2h, row4h)));
 }
-#endif  // CRYPTOPP_BOOL_ARM32 && CRYPTOPP_ARM_NEON_AVAILABLE
+#endif  // CRYPTOPP_ARM_NEON_AVAILABLE
 
 NAMESPACE_END

blake2.cpp

@@ -17,6 +17,12 @@ NAMESPACE_BEGIN(CryptoPP)
 // #undef CRYPTOPP_SSE41_AVAILABLE
 // #undef CRYPTOPP_ARM_NEON_AVAILABLE
 
+// Disable NEON/ASIMD for Cortex-A53 and A57. The shifts are too slow and C/C++ is about
+// 3 cpb faster than NEON/ASIMD. Also see http://github.com/weidai11/cryptopp/issues/367.
+#if (defined(__aarch32__) || defined(__aarch64__)) && defined(CRYPTOPP_SLOW_ARMV8_SHIFT)
+# undef CRYPTOPP_ARM_NEON_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);
 
@@ -25,8 +31,7 @@ extern void BLAKE2_Compress32_SSE4(const byte* input, BLAKE2_State<word32, false
 extern void BLAKE2_Compress64_SSE4(const byte* input, BLAKE2_State<word64, true>& state);
 #endif
 
-// Disable NEON for Cortex-A53 and A57. Also see http://github.com/weidai11/cryptopp/issues/367
+#if CRYPTOPP_ARM_NEON_AVAILABLE
-#if CRYPTOPP_BOOL_ARM32 && CRYPTOPP_ARM_NEON_AVAILABLE
 extern void BLAKE2_Compress32_NEON(const byte* input, BLAKE2_State<word32, false>& state);
 extern void BLAKE2_Compress64_NEON(const byte* input, BLAKE2_State<word64, true>& state);
 #endif
@@ -93,7 +98,7 @@ pfnCompress64 InitializeCompress64Fn()
 #if CRYPTOPP_SSE41_AVAILABLE
 		HasSSE41() ? &BLAKE2_Compress64_SSE4 :
 #endif
-#if CRYPTOPP_BOOL_ARM32 && CRYPTOPP_ARM_NEON_AVAILABLE
+#if CRYPTOPP_ARM_NEON_AVAILABLE
 		HasNEON() ? &BLAKE2_Compress64_NEON :
 #endif
 		&BLAKE2_Compress64_CXX;
@@ -105,7 +110,7 @@ pfnCompress32 InitializeCompress32Fn()
 #if CRYPTOPP_SSE41_AVAILABLE
 		HasSSE41() ? &BLAKE2_Compress32_SSE4 :
 #endif
-#if CRYPTOPP_BOOL_ARM32 && CRYPTOPP_ARM_NEON_AVAILABLE
+#if CRYPTOPP_ARM_NEON_AVAILABLE
 		HasNEON() ? &BLAKE2_Compress32_NEON :
 #endif
 		&BLAKE2_Compress32_CXX;

config.h

@@ -118,6 +118,12 @@
 // of 'b', 'o', 'h' or '.' (the last for decimal).
 // #define CRYPTOPP_USE_STD_SHOWBASE
 
+// Define this if ARMv8 shifts are slow. ARM Cortex-A53 and Cortex-A57 shift
+// operation perform poorly, so NEON and ASIMD code that relies on shifts
+// or rotates often performs worse than regular C/C++ code. Also see
+// http://github.com/weidai11/cryptopp/issues/367.
+#define CRYPTOPP_SLOW_ARMV8_SHIFT 1
+
 // Define this if you want to decouple AlgorithmParameters and Integer
 // The decoupling should make it easier for the linker to remove Integer
 // related code for those who do not need Integer, and avoid a potential