Cleanup ARIA SSE and NEON code

2 files changed, 82 insertions, 46 deletions

diff --git a/aria-simd.cpp b/aria-simd.cpp
index ce16ada5..9c741d7e 100644
--- a/aria-simd.cpp
+++ b/aria-simd.cpp
@@ -29,10 +29,6 @@
 #define M128_CAST(x) ((__m128i *)(void *)(x))

 #define CONST_M128_CAST(x) ((const __m128i *)(const void *)(x))

 

-// GCC cast warning

-#define UINT32_CAST(x) ((uint32_t *)(void *)(x))

-#define CONST_UINT32_CAST(x) ((const uint32_t *)(const void *)(x))

-

 NAMESPACE_BEGIN(CryptoPP)

 NAMESPACE_BEGIN(ARIATab)

 

@@ -45,6 +41,17 @@ extern const word32 KRK[3][4];
 NAMESPACE_END

 NAMESPACE_END

 

+ANONYMOUS_NAMESPACE_BEGIN

+

+using CryptoPP::byte;

+using CryptoPP::word32;

+

+inline byte ARIA_BRF(const word32 x, const int y) {

+	return GETBYTE(x, y);

+}

+

+ANONYMOUS_NAMESPACE_END

+

 NAMESPACE_BEGIN(CryptoPP)

 

 using CryptoPP::ARIATab::S1;

@@ -58,22 +65,23 @@ using CryptoPP::ARIATab::KRK;
 template <unsigned int N>

 inline void ARIA_GSRK_NEON(const uint32x4_t X, const uint32x4_t Y, byte RK[16])

 {

-	static const unsigned int Q1 = (4-(N/32)) % 4;

-	static const unsigned int Q2 = (3-(N/32)) % 4;

-	static const unsigned int R = N % 32;

+	enum { Q1 = (4-(N/32)) % 4,

+	       Q2 = (3-(N/32)) % 4,

+	       R = N % 32

+	};

 

-	vst1q_u32(UINT32_CAST(RK),

+	vst1q_u8(RK, vreinterpretq_u8_u32(

 		veorq_u32(X, veorq_u32(

 			vshrq_n_u32(vextq_u32(Y, Y, Q1), R),

-			vshlq_n_u32(vextq_u32(Y, Y, Q2), 32-R))));

+			vshlq_n_u32(vextq_u32(Y, Y, Q2), 32-R)))));

 }

 

 void ARIA_UncheckedSetKey_Schedule_NEON(byte* rk, word32* ws, unsigned int keylen)

 {

-	const uint32x4_t w0 = vld1q_u32(CONST_UINT32_CAST(ws+ 0));

-	const uint32x4_t w1 = vld1q_u32(CONST_UINT32_CAST(ws+ 8));

-	const uint32x4_t w2 = vld1q_u32(CONST_UINT32_CAST(ws+12));

-	const uint32x4_t w3 = vld1q_u32(CONST_UINT32_CAST(ws+16));

+	const uint32x4_t w0 = vld1q_u32(ws+ 0);

+	const uint32x4_t w1 = vld1q_u32(ws+ 8);

+	const uint32x4_t w2 = vld1q_u32(ws+12);

+	const uint32x4_t w3 = vld1q_u32(ws+16);

 

 	ARIA_GSRK_NEON<19>(w0, w1, rk +   0);

 	ARIA_GSRK_NEON<19>(w1, w2, rk +  16);

@@ -102,22 +110,49 @@ void ARIA_UncheckedSetKey_Schedule_NEON(byte* rk, word32* ws, unsigned int keyle
 	}

 }

 

-void ARIA_ProcessAndXorBlock_Xor_NEON(const byte* xorBlock, byte* outBlock)

+void ARIA_ProcessAndXorBlock_NEON(const byte* xorBlock, byte* outBlock, const byte *rk, word32 *t)

 {

-	vst1q_u32(UINT32_CAST(outBlock), veorq_u32(

-		vld1q_u32(CONST_UINT32_CAST(outBlock)),

-		vld1q_u32(CONST_UINT32_CAST(xorBlock))));

+	outBlock[ 0] = (byte)(X1[ARIA_BRF(t[0],3)]   );

+	outBlock[ 1] = (byte)(X2[ARIA_BRF(t[0],2)]>>8);

+	outBlock[ 2] = (byte)(S1[ARIA_BRF(t[0],1)]   );

+	outBlock[ 3] = (byte)(S2[ARIA_BRF(t[0],0)]   );

+	outBlock[ 4] = (byte)(X1[ARIA_BRF(t[1],3)]   );

+	outBlock[ 5] = (byte)(X2[ARIA_BRF(t[1],2)]>>8);

+	outBlock[ 6] = (byte)(S1[ARIA_BRF(t[1],1)]   );

+	outBlock[ 7] = (byte)(S2[ARIA_BRF(t[1],0)]   );

+	outBlock[ 8] = (byte)(X1[ARIA_BRF(t[2],3)]   );

+	outBlock[ 9] = (byte)(X2[ARIA_BRF(t[2],2)]>>8);

+	outBlock[10] = (byte)(S1[ARIA_BRF(t[2],1)]   );

+	outBlock[11] = (byte)(S2[ARIA_BRF(t[2],0)]   );

+	outBlock[12] = (byte)(X1[ARIA_BRF(t[3],3)]   );

+	outBlock[13] = (byte)(X2[ARIA_BRF(t[3],2)]>>8);

+	outBlock[14] = (byte)(S1[ARIA_BRF(t[3],1)]   );

+	outBlock[15] = (byte)(S2[ARIA_BRF(t[3],0)]   );

+

+	// 'outBlock' and 'xorBlock' may be unaligned.

+	if (xorBlock != NULLPTR)

+	{

+		vst1q_u8(outBlock,

+			veorq_u8(

+				vld1q_u8(xorBlock),

+				veorq_u8(

+					vld1q_u8(outBlock),

+					vrev32q_u8(vld1q_u8((rk))))));

+	}

+	else

+	{

+		vst1q_u8(outBlock,

+			veorq_u8(

+				vld1q_u8(outBlock),

+				vrev32q_u8(vld1q_u8(rk))));

+	}

 }

 

 #endif  // CRYPTOPP_ARM_NEON_AVAILABLE

 

 #if (CRYPTOPP_SSSE3_AVAILABLE)

 

-inline byte ARIA_BRF(const word32 x, const int y) {

-	return GETBYTE(x, y);

-}

-

-void ARIA_ProcessAndXorBlock_Xor_SSSE3(const byte* xorBlock, byte* outBlock, const byte *rk, word32 *t)

+void ARIA_ProcessAndXorBlock_SSSE3(const byte* xorBlock, byte* outBlock, const byte *rk, word32 *t)

 {

 	const __m128i MASK = _mm_set_epi8(12,13,14,15, 8,9,10,11, 4,5,6,7, 0,1,2,3);

 

@@ -138,18 +173,22 @@ void ARIA_ProcessAndXorBlock_Xor_SSSE3(const byte* xorBlock, byte* outBlock, con
 	outBlock[14] = (byte)(S1[ARIA_BRF(t[3],1)]   );

 	outBlock[15] = (byte)(S2[ARIA_BRF(t[3],0)]   );

 

-	// 'outBlock' may be unaligned.

-	_mm_storeu_si128(M128_CAST(outBlock),

-		_mm_xor_si128(_mm_loadu_si128(CONST_M128_CAST(outBlock)),

-			_mm_shuffle_epi8(_mm_load_si128(CONST_M128_CAST(rk)), MASK)));

-

 	// 'outBlock' and 'xorBlock' may be unaligned.

 	if (xorBlock != NULLPTR)

 	{

 		_mm_storeu_si128(M128_CAST(outBlock),

 			_mm_xor_si128(

-				_mm_loadu_si128(CONST_M128_CAST(outBlock)),

-				_mm_loadu_si128(CONST_M128_CAST(xorBlock))));

+				_mm_loadu_si128(CONST_M128_CAST(xorBlock)),

+				_mm_xor_si128(

+					_mm_loadu_si128(CONST_M128_CAST(outBlock)),

+					_mm_shuffle_epi8(_mm_load_si128(CONST_M128_CAST(rk)), MASK)))

+			);

+	}

+	else

+	{

+		_mm_storeu_si128(M128_CAST(outBlock),

+			_mm_xor_si128(_mm_loadu_si128(CONST_M128_CAST(outBlock)),

+				_mm_shuffle_epi8(_mm_load_si128(CONST_M128_CAST(rk)), MASK)));

 	}

 }

 

diff --git a/aria.cpp b/aria.cpp
index 52cc2eb2..8f3f221a 100644
--- a/aria.cpp
+++ b/aria.cpp
@@ -85,11 +85,11 @@ inline byte ARIA_BRF(const word32 x, const int y) {
 

 #if (CRYPTOPP_ARM_NEON_AVAILABLE)

 extern void ARIA_UncheckedSetKey_Schedule_NEON(byte* rk, word32* ws, unsigned int keylen);

-extern void ARIA_ProcessAndXorBlock_Xor_NEON(const byte* xorBlock, byte* outblock);

+extern void ARIA_ProcessAndXorBlock_NEON(const byte* xorBlock, byte* outblock, const byte *rk, word32 *t);

 #endif

 

 #if (CRYPTOPP_SSSE3_AVAILABLE)

-extern void ARIA_ProcessAndXorBlock_Xor_SSSE3(const byte* xorBlock, byte* outBlock, const byte *rk, word32 *t);

+extern void ARIA_ProcessAndXorBlock_SSSE3(const byte* xorBlock, byte* outBlock, const byte *rk, word32 *t);

 #endif

 

 // n-bit right shift of Y XORed to X

@@ -283,12 +283,19 @@ void ARIA::Base::ProcessAndXorBlock(const byte *inBlock, const byte *xorBlock, b
 #if CRYPTOPP_ENABLE_ARIA_SSSE3_INTRINSICS

 	if (HasSSSE3())

 	{

-		ARIA_ProcessAndXorBlock_Xor_SSSE3(xorBlock, outBlock, rk, t);

+		ARIA_ProcessAndXorBlock_SSSE3(xorBlock, outBlock, rk, t);

 		return;

 	}

 	else

 #endif  // CRYPTOPP_ENABLE_ARIA_SSSE3_INTRINSICS

-

+#if (CRYPTOPP_ARM_NEON_AVAILABLE)

+	if (HasNEON())

+	{

+		ARIA_ProcessAndXorBlock_NEON(xorBlock, outBlock, rk, t);

+		return;

+	}

+	else

+#endif  // CRYPTOPP_ARM_NEON_AVAILABLE

 #if (CRYPTOPP_LITTLE_ENDIAN)

 	{

 		outBlock[ 0] = (byte)(X1[ARIA_BRF(t[0],3)]   ) ^ rk[ 3];

@@ -329,19 +336,9 @@ void ARIA::Base::ProcessAndXorBlock(const byte *inBlock, const byte *xorBlock, b
 	}

 #endif  // CRYPTOPP_LITTLE_ENDIAN

 

-#if CRYPTOPP_ARM_NEON_AVAILABLE

-	if (HasNEON())

-	{

-		if (xorBlock != NULLPTR)

-			ARIA_ProcessAndXorBlock_Xor_NEON(xorBlock, outBlock);

-	}

-	else

-#endif  // CRYPTOPP_ARM_NEON_AVAILABLE

-	{

-		if (xorBlock != NULLPTR)

-			for (unsigned int n=0; n<ARIA::BLOCKSIZE; ++n)

-				outBlock[n] ^= xorBlock[n];

-	}

+	if (xorBlock != NULLPTR)

+		for (unsigned int n=0; n<ARIA::BLOCKSIZE; ++n)

+			outBlock[n] ^= xorBlock[n];

 }

 

 NAMESPACE_END