Enable Power7 for Simon and Speck (PR #909)

1 files changed, 70 insertions, 14 deletions

diff --git a/simon.cpp b/simon.cpp
index 91f8bf1d..a35762de 100644
--- a/simon.cpp
+++ b/simon.cpp
@@ -212,7 +212,7 @@ extern size_t SIMON128_Dec_AdvancedProcessBlocks_NEON(const word64* subKeys, siz
     const byte *inBlocks, const byte *xorBlocks, byte *outBlocks, size_t length, word32 flags);

 #endif

 

-#if defined(CRYPTOPP_SSE41_AVAILABLE)

+#if (CRYPTOPP_SSE41_AVAILABLE)

 extern size_t SIMON64_Enc_AdvancedProcessBlocks_SSE41(const word32* subKeys, size_t rounds,

     const byte *inBlocks, const byte *xorBlocks, byte *outBlocks, size_t length, word32 flags);

 

@@ -220,7 +220,7 @@ extern size_t SIMON64_Dec_AdvancedProcessBlocks_SSE41(const word32* subKeys, siz
     const byte *inBlocks, const byte *xorBlocks, byte *outBlocks, size_t length, word32 flags);

 #endif

 

-#if defined(CRYPTOPP_SSSE3_AVAILABLE)

+#if (CRYPTOPP_SSSE3_AVAILABLE)

 extern size_t SIMON128_Enc_AdvancedProcessBlocks_SSSE3(const word64* subKeys, size_t rounds,

     const byte *inBlocks, const byte *xorBlocks, byte *outBlocks, size_t length, word32 flags);

 

@@ -228,7 +228,13 @@ extern size_t SIMON128_Dec_AdvancedProcessBlocks_SSSE3(const word64* subKeys, si
     const byte *inBlocks, const byte *xorBlocks, byte *outBlocks, size_t length, word32 flags);

 #endif

 

-#if (CRYPTOPP_ALTIVEC_AVAILABLE)

+#if (CRYPTOPP_POWER7_AVAILABLE)

+extern size_t SIMON64_Enc_AdvancedProcessBlocks_POWER7(const word32* subKeys, size_t rounds,

+    const byte *inBlocks, const byte *xorBlocks, byte *outBlocks, size_t length, word32 flags);

+

+extern size_t SIMON64_Dec_AdvancedProcessBlocks_POWER7(const word32* subKeys, size_t rounds,

+    const byte *inBlocks, const byte *xorBlocks, byte *outBlocks, size_t length, word32 flags);

+#elif (CRYPTOPP_ALTIVEC_AVAILABLE)

 extern size_t SIMON64_Enc_AdvancedProcessBlocks_ALTIVEC(const word32* subKeys, size_t rounds,

     const byte *inBlocks, const byte *xorBlocks, byte *outBlocks, size_t length, word32 flags);

 

@@ -255,9 +261,9 @@ std::string SIMON64::Base::AlgorithmProvider() const
     if (HasNEON())

         return "NEON";

 # endif

-# if (CRYPTOPP_POWER8_AVAILABLE)

-    if (HasPower8())

-        return "Power8";

+# if (CRYPTOPP_POWER7_AVAILABLE)

+    if (HasPower7())

+        return "Power7";

 # endif

 # if (CRYPTOPP_ALTIVEC_AVAILABLE)

     if (HasAltivec())

@@ -267,6 +273,29 @@ std::string SIMON64::Base::AlgorithmProvider() const
     return "C++";

 }

 

+unsigned int SIMON64::Base::OptimalDataAlignment() const

+{

+#if (CRYPTOPP_SIMON64_ADVANCED_PROCESS_BLOCKS)

+# if (CRYPTOPP_SSE41_AVAILABLE)

+    if (HasSSE41())

+        return 16;

+# endif

+# if (CRYPTOPP_ARM_NEON_AVAILABLE)

+    if (HasNEON())

+        return 4;

+# endif

+# if (CRYPTOPP_POWER7_AVAILABLE)

+    if (HasPower7())

+        return 4;

+# endif

+# if (CRYPTOPP_ALTIVEC_AVAILABLE)

+    if (HasAltivec())

+        return 16;

+# endif

+#endif

+    return GetAlignmentOf<word32>();

+}

+

 void SIMON64::Base::UncheckedSetKey(const byte *userKey, unsigned int keyLength, const NameValuePairs &params)

 {

     CRYPTOPP_ASSERT(keyLength == 12 || keyLength == 16);

@@ -371,6 +400,25 @@ std::string SIMON128::Base::AlgorithmProvider() const
     return "C++";

 }

 

+unsigned int SIMON128::Base::OptimalDataAlignment() const

+{

+#if (CRYPTOPP_SIMON128_ADVANCED_PROCESS_BLOCKS)

+# if (CRYPTOPP_SSSE3_AVAILABLE)

+    if (HasSSSE3())

+        return 16;

+# endif

+# if (CRYPTOPP_ARM_NEON_AVAILABLE)

+    if (HasNEON())

+        return 8;

+# endif

+# if (CRYPTOPP_POWER8_AVAILABLE)

+    if (HasPower8())

+        return 8;

+# endif

+#endif

+    return GetAlignmentOf<word64>();

+}

+

 void SIMON128::Base::UncheckedSetKey(const byte *userKey, unsigned int keyLength, const NameValuePairs &params)

 {

     CRYPTOPP_ASSERT(keyLength == 16 || keyLength == 24 || keyLength == 32);

@@ -459,11 +507,11 @@ void SIMON128::Dec::ProcessAndXorBlock(const byte *inBlock, const byte *xorBlock
     OutBlock oblk(xorBlock, outBlock); oblk(m_wspace[3])(m_wspace[2]);

 }

 

-#if defined(CRYPTOPP_SIMON64_ADVANCED_PROCESS_BLOCKS)

+#if (CRYPTOPP_SIMON64_ADVANCED_PROCESS_BLOCKS)

 size_t SIMON64::Enc::AdvancedProcessBlocks(const byte *inBlocks, const byte *xorBlocks,

         byte *outBlocks, size_t length, word32 flags) const

 {

-#if defined(CRYPTOPP_SSE41_AVAILABLE)

+#if (CRYPTOPP_SSE41_AVAILABLE)

     if (HasSSE41())

         return SIMON64_Enc_AdvancedProcessBlocks_SSE41(m_rkeys, (size_t)m_rounds,

             inBlocks, xorBlocks, outBlocks, length, flags);

@@ -473,7 +521,11 @@ size_t SIMON64::Enc::AdvancedProcessBlocks(const byte *inBlocks, const byte *xor
         return SIMON64_Enc_AdvancedProcessBlocks_NEON(m_rkeys, (size_t)m_rounds,

             inBlocks, xorBlocks, outBlocks, length, flags);

 #endif

-#if (CRYPTOPP_ALTIVEC_AVAILABLE)

+#if (CRYPTOPP_POWER7_AVAILABLE)

+    if (HasPower7())

+        return SIMON64_Enc_AdvancedProcessBlocks_POWER7(m_rkeys, (size_t)m_rounds,

+            inBlocks, xorBlocks, outBlocks, length, flags);

+#elif (CRYPTOPP_ALTIVEC_AVAILABLE)

     if (HasAltivec())

         return SIMON64_Enc_AdvancedProcessBlocks_ALTIVEC(m_rkeys, (size_t)m_rounds,

             inBlocks, xorBlocks, outBlocks, length, flags);

@@ -484,7 +536,7 @@ size_t SIMON64::Enc::AdvancedProcessBlocks(const byte *inBlocks, const byte *xor
 size_t SIMON64::Dec::AdvancedProcessBlocks(const byte *inBlocks, const byte *xorBlocks,

         byte *outBlocks, size_t length, word32 flags) const

 {

-#if defined(CRYPTOPP_SSE41_AVAILABLE)

+#if (CRYPTOPP_SSE41_AVAILABLE)

     if (HasSSE41())

         return SIMON64_Dec_AdvancedProcessBlocks_SSE41(m_rkeys, (size_t)m_rounds,

             inBlocks, xorBlocks, outBlocks, length, flags);

@@ -494,7 +546,11 @@ size_t SIMON64::Dec::AdvancedProcessBlocks(const byte *inBlocks, const byte *xor
         return SIMON64_Dec_AdvancedProcessBlocks_NEON(m_rkeys, (size_t)m_rounds,

             inBlocks, xorBlocks, outBlocks, length, flags);

 #endif

-#if (CRYPTOPP_ALTIVEC_AVAILABLE)

+#if (CRYPTOPP_POWER7_AVAILABLE)

+    if (HasPower7())

+        return SIMON64_Dec_AdvancedProcessBlocks_POWER7(m_rkeys, (size_t)m_rounds,

+            inBlocks, xorBlocks, outBlocks, length, flags);

+#elif (CRYPTOPP_ALTIVEC_AVAILABLE)

     if (HasAltivec())

         return SIMON64_Dec_AdvancedProcessBlocks_ALTIVEC(m_rkeys, (size_t)m_rounds,

             inBlocks, xorBlocks, outBlocks, length, flags);

@@ -503,11 +559,11 @@ size_t SIMON64::Dec::AdvancedProcessBlocks(const byte *inBlocks, const byte *xor
 }

 #endif  // CRYPTOPP_SIMON64_ADVANCED_PROCESS_BLOCKS

 

-#if defined(CRYPTOPP_SIMON128_ADVANCED_PROCESS_BLOCKS)

+#if (CRYPTOPP_SIMON128_ADVANCED_PROCESS_BLOCKS)

 size_t SIMON128::Enc::AdvancedProcessBlocks(const byte *inBlocks, const byte *xorBlocks,

         byte *outBlocks, size_t length, word32 flags) const

 {

-#if defined(CRYPTOPP_SSSE3_AVAILABLE)

+#if (CRYPTOPP_SSSE3_AVAILABLE)

     if (HasSSSE3())

         return SIMON128_Enc_AdvancedProcessBlocks_SSSE3(m_rkeys, (size_t)m_rounds,

             inBlocks, xorBlocks, outBlocks, length, flags);

@@ -528,7 +584,7 @@ size_t SIMON128::Enc::AdvancedProcessBlocks(const byte *inBlocks, const byte *xo
 size_t SIMON128::Dec::AdvancedProcessBlocks(const byte *inBlocks, const byte *xorBlocks,

         byte *outBlocks, size_t length, word32 flags) const

 {

-#if defined(CRYPTOPP_SSSE3_AVAILABLE)

+#if (CRYPTOPP_SSSE3_AVAILABLE)

     if (HasSSSE3())

         return SIMON128_Dec_AdvancedProcessBlocks_SSSE3(m_rkeys, (size_t)m_rounds,

             inBlocks, xorBlocks, outBlocks, length, flags);